# ALZHEIMER'S DISEASE AND THE FORNIX

EDITED BY: Kenichi Oishi and Constantine G. Lyketsos PUBLISHED IN: Frontiers in Aging Neuroscience

#### *Frontiers Copyright Statement*

*© Copyright 2007-2016 Frontiers Media SA. All rights reserved. All content included on this site, such as text, graphics, logos, button icons, images, video/audio clips, downloads, data compilations and software, is the property of or is licensed to Frontiers Media SA ("Frontiers") or its licensees and/or subcontractors. The copyright in the text of individual articles is the property of their respective authors, subject to a license granted to Frontiers.*

*The compilation of articles constituting this e-book, wherever published, as well as the compilation of all other content on this site, is the exclusive property of Frontiers. For the conditions for downloading and copying of e-books from Frontiers' website, please see the Terms for Website Use. If purchasing Frontiers e-books from other websites or sources, the conditions of the website concerned apply.*

*Images and graphics not forming part of user-contributed materials may not be downloaded or copied without permission.*

*Individual articles may be downloaded and reproduced in accordance with the principles of the CC-BY licence subject to any copyright or other notices. They may not be re-sold as an e-book.*

*As author or other contributor you grant a CC-BY licence to others to reproduce your articles, including any graphics and third-party materials supplied by you, in accordance with the Conditions for Website Use and subject to any copyright notices which you include in connection with your articles and materials.*

> *All copyright, and all rights therein, are protected by national and international copyright laws.*

*The above represents a summary only. For the full conditions see the Conditions for Authors and the Conditions for Website Use.*

ISSN 1664-8714 ISBN 978-2-88919-959-4 DOI 10.3389/978-2-88919-959-4

# About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

# Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

# Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

# What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# **ALZHEIMER'S DISEASE AND THE FORNIX**

Topic Editors: **Kenichi Oishi,** Johns Hopkins University, USA **Constantine G. Lyketsos,** Johns Hopkins University, USA

Coronal view of the human brain. Image by Kenichi Oishi

This e-book focuses primarily on the role of the fornix as a functional, prognostic, and diagnostic marker of Alzheimer's disease (AD), and the application of such a marker in clinical practice.

Researchers have long been focused on the cortical pathology of AD, since the most important pathologic features are the senile plaques found in the cortex, and the neurofibrillary tangles and neuronal loss that start from the entorhinal cortex and the hippocampus. In addition to gray matter structures, histopathological studies indicate that the white matter is also altered in AD. The fornix is a white matter bundle that constitutes a core element of the limbic circuits, and is one of the most important anatomical structures related to memory. The fornices originate from the bilateral hippocampi, merge at the midline of the brain, again divide into the left and right side, and then into the precommissural and the postcommissural fibers, and terminate at the septal nuclei, nucleus accumbens (precommissural fornix), and

hypothalamus (postcommissural fornix). These functional and anatomical features of the fornix have naturally captured researchers' attention as possible diagnostic and prognostic markers of AD. Growing evidence indicates that the alterations seen in the fornix are potentially a good marker with which to predict future conversion from mild cognitive impairment to AD, and even from a cognitively normal state to AD. The degree of alteration is correlated with the degree of memory impairment, indicating the potential for the use of the fornix as a functional marker. Moreover, there have been attempts to stimulate the fornix to recover the cognitive function lost with AD. Our goal is to provide information about the status of current research and to facilitate further scientific and clinical advancement in this topic.

**Citation:** Oishi, K., Lyketsos, C. G., eds. (2016). Alzheimer's Disease and the Fornix. Lausanne: Frontiers Media. doi: 10.3389/978-2-88919-959-4

# Table of Contents


# **Chapter 1: Normal anatomy and development**


Qiaowen Yu, Yun Peng, Virendra Mishra, Austin Ouyang, Hang Li, Hong Zhang, Min Chen, Shuwei Liu and Hao Huang

# **Chapter 2: Changes in anatomy, connectivity, and function**


Elizabeth G. Kehoe, Dervla Farrell, Claudia Metzler-Baddeley, Brian A. Lawlor, Rose Anne Kenny, Declan Lyons, Jonathan P. McNulty, Paul G. Mullins, Damien Coyle and Arun L. Bokde

*55 Correlations between limbic white matter and cognitive function in temporallobe epilepsy, preliminary findings*

Ryan P. D. Alexander, Luis Concha, Thomas J. Snyder, Christian Beaulieu and Donald William Gross

# **Chapter 3: Detection of the fornix degeneration and methodological aspects**

*61 Fractional anisotropy of the fornix and hippocampal atrophy in Alzheimer's disease*

Kejal Kantarci


# *81 Diffusion tensor imaging in Alzheimer's disease: insights into the limbicdiencephalic network and methodological considerations*

Julio Acosta-Cabronero and Peter J. Nestor

# **Chapter 4: Toward clinical application**

# *102 Alzheimer's disease and the fornix*

Kenichi Oishi and Constantine G. Lyketsos

# Editorial: Alzheimer's Disease and the Fornix

#### Kenichi Oishi <sup>1</sup> \* and Constantine G. Lyketsos <sup>2</sup>

<sup>1</sup> The Russell H. Morgan Department of Radiology and Radiological Sciences, Johns Hopkins University, Baltimore, MD, USA, <sup>2</sup> Department of Psychiatry and Behavioral Sciences, Johns Hopkins University, Baltimore, MD, USA

Keywords: Alzheimer's disease, fornix, limbic, diffusion tensor imaging, normal aging, memory, cognition, mild cognitive impairment

#### **The Editorial on the Research Topic**

#### **Alzheimer's Disease and the Fornix**

The fornix is a white matter bundle that connects the hippocampus with other limbic structures. It appears in the literature as early as 1543 in a historical publication of De Humani Corporis Fabrica by Andreas Vesalius (Swanson, 2014). The fornix is important for episodic memory recall (Tsivilis et al., 2008), which is impaired in Alzheimer's disease (AD). Alterations in the fornix were first observed in post-mortem AD brains (Hopper and Vogel, 1976). This volume focuses on the role of the fornix, and other limbic fibers, in the disease mechanisms of AD with some attention to how this might be applied in clinical practice.

The observation of limbic fibers in vivo forms a basis for understanding normal anatomy and alterations caused by various diseases. Mori and Aggarwal were able to observe the fornix, cingulum, and stria terminalis in mice and humans, using T1-weighted and diffusion tensor imaging.

In adult human brains, the limbic fibers are known to connect the structures of the default mode network (DMN), but the development of these fibers is less well understood. Yu et al. demonstrated that the developmental curve of DTI-derived measures of fornix integrity appear logarithmic, with rapid changes until 2 years of age followed by slow changes until 25 years. Development of the cingulate cingulum is disproportionally rapid during this period, but development of the hippocampal cingulum and the fornix is proportional. Notably, the functional and anatomic connectivity of the DMN is already established in the early postnatal period.

#### Edited and reviewed by:

Rodrigo Orlando Kuljiš, University of Miami School of Medicine, USA

#### \*Correspondence:

Kenichi Oishi koishi@mri.jhu.edu

Received: 01 April 2016 Accepted: 10 June 2016 Published: 23 June 2016

#### Citation:

Oishi K and Lyketsos CG (2016) Editorial: Alzheimer's Disease and the Fornix. Front. Aging Neurosci. 8:149. doi: 10.3389/fnagi.2016.00149

The fornix is among the white matter structures that mature early during development. Douet and Chang reviewed changes in DTI measures in the fornix during development and aging. Development of the fornix peaks in late adolescence, followed by pruning and then degeneration. Fractional anisotropy (FA) values correlate with cognitive performance in various age groups, including children, young adults, and the elderly. Correlations are seen in various brain diseases including schizophrenia, multiple sclerosis, Parkinson's, and epilepsy.

Normal aging affects the anatomy of the fornix. Kantarci proposed a hippocampus-fornix axis in which microstructural alterations in both hippocampus and fornix affect each other. In AD, it is likely that neuronal damage in hippocampus and axonal damage in fornix affect each other, but alterations in the fornix in normal aging are likely the consequence of age-related, non-specific axonal and myelin damage.

Less is known about the relationship between alterations in fornix and functional connectivity. Kehoe et al. investigated whether diffusivity in fornix is related to functional connectivity between thalamus and hippocampus. Several diffusivity measures were correlated with functional connectivity among cognitively normal elderly, but this correlation was not seen in individuals with amnestic mild cognitive impairment (MCI). This suggests that the pathological processes of amnestic MCI mitigate the structural-functional relationship that is normally seen.

Tract-based spatial statistics (TBSS) are commonly used to analyze neuroanatomical alterations related to AD. Acosta-Cabronero and Nestor reviewed AD studies that applied TBSS. In early AD, increases in the first eigenvalue were identified in fornix, parahippocampal white matter, and anterior thalamus. The authors emphasized the importance of technological factors that affect results of clinical DTI studies. These include the basis of diffusion-weighted signals and tensor calculation, the imaging parameters, the post-processing methodology, the subject cohort, multi-center study designs, and inclusion criteria. This review is quite helpful to investigators who plan to study DTI of AD or other diseases.

Nowrangi and Rosenberg summarized DTI-derived scalar measures as correlates of cognitive functions or Aβ deposition in AD, as predictors of future conversion from MCI to AD, and as potential targets for deep brain stimulation. Fletcher et al. applied these markers to a cohort of cognitively normal elderly to predict later occurrence of brain atrophy. Reduced FA was used as a marker for white matter alterations in various structures, including fornix, genu, splenium of the corpus callosum, and anterior and posterior cingulum. Volume reduction was used to evaluate neurodegeneration in the hippocampus and in the ventral and dorsal entorhinal cortices. Fornix FA was the most sensitive marker among structures investigated in predicting future brain atrophy typically seen in AD.

One direction for the future in neuroimaging studies is the application of research to the clinical arena, in which prediction is an important theme. Oishi and Lyketsos reviewed DTI analysis methods reported to detect anatomical abnormalities in the AD brain, especially fornix, and discussed the potential for the early diagnosis, prediction of cognitive worsening, and therapeutic targets.

# REFERENCES

Hopper, M. W., and Vogel, F. S. (1976). The limbic system in Alzheimer's disease. A neuropathologic investigation. Am. J. Pathol. 85, 1–20.


Disease specificity is often an issue in clinical image reading. Although alterations in the fornix are often seen in AD, such alterations are also seen in other diseases, such as temporal lobe epilepsy. Alexander et al. investigated the relationship between limbic fiber integrity and cognitive function using DTI in patients with temporal lobe epilepsy. They identified a correlation between FA of the left fornix and processing speed, but not between T2 of the hippocampus and processing speed. This suggested that the relation between fornix injury and functional decline is not disease-specific, but rather, the result of injury in a neuronal network with specific neuronal functions.

In summary, the 10 articles included in this volume cover anatomy, development, aging, disease, and functional correlations or clinical significance, which are informative for readers who plan to investigate the fornix in AD or other diseases.

# AUTHOR CONTRIBUTIONS

KO wrote preliminary draft and CL finalized the manuscript.

# ACKNOWLEDGMENTS

This publication was made possible by grants P50AG005146 and R01AG042165 from the National Institutes of Health, and in Health Pilot Project grant from the Johns Hopkins Individualized Health Initiative. The contents of this paper are solely the responsibility of the authors and do not necessarily represent the official view of NIH or the Johns Hopkins Individualized Health Initiative. We thank Ms. Mary McAllister for her manuscript editing.

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2016 Oishi and Lyketsos. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# In vivo magnetic resonance imaging of the human limbic white matter

# **Susumu Mori 1,2\* and Manisha Aggarwal <sup>1</sup>**

<sup>1</sup> Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, USA <sup>2</sup> F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, USA

#### **Edited by:**

P. Hemachandra Reddy, Texas Tech University, USA

#### **Reviewed by:**

Koteswara Rao Valasani, The University of Kansas, USA Ramesh Kandimalla, Texas Tech University, USA

#### **\*Correspondence:**

Susumu Mori, Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, 330 Traylor Building, 720 Rutland Avenue, Baltimore, MD 21205, USA e-mail: susumu@mri.jhu.edu

The limbic system mediates memory, behavior, and emotional output in the human brain, and is implicated in the pathology of Alzheimer's disease and a wide spectrum of related neurological disorders. In vivo magnetic resonance imaging (MRI) of structural components comprising the limbic system and their interconnections via white matter pathways in the human brain has helped define current understanding of the limbic model based on the classical circuit proposed by Papez. MRI techniques, including diffusion MR imaging, provide a non-invasive method to characterize white matter tracts of the limbic system, and investigate pathological changes that affect these pathways in clinical settings. This review focuses on delineation of the anatomy of major limbic tracts in the human brain, namely, the cingulum, the fornix and fimbria, and the stria terminalis, based on in vivo MRI contrasts. The detailed morphology and intricate trajectories of these pathways that can be identified using relaxometry-based and diffusion-weighted MRI provide an important anatomical reference for evaluation of clinical disorders commonly associated with limbic pathology.

**Keywords: limbic system, MRI, diffusion, human, white matter, in vivo**

# **INTRODUCTION**

The limbic system consists of a group of interconnected nuclei and cortical structures in the brain that mediate emotion, memory, and behavior (Patestas and Gartner, 2006). The classical circuit described by Papez includes important white matter pathways interlinking the hippocampus, mammillary bodies, anterior thalamic nuclei, cingulate gyrus (Cg), and the parahippocampal gyrus, that form a closed loop in each hemisphere (Papez, 1995). The complex connections of the limbic system are implicated in a wide array of neurological disorders including Alzheimer's disease, mild cognitive impairment, temporal lobe epilepsy, and schizophrenia, and continue to be defined using functional neuroimaging and magnetic resonance imaging (MRI) studies (Naidich et al., 1987a,b; Mark et al., 1995; Concha et al., 2010; Catani et al., 2013). This purpose of this review is to focus on the use of MRI methods for delineation of some of the major white matter tracts that are associated with the limbic system.

White matter tracts in the human brain can be broadly classified into three categories; association, projection, and commissural tracts (Nieuwenhuys et al., 2008). Association tracts establish connections across different cortical regions within the same hemisphere. Short association fibers form connections between different gyri within the same lobe, while long association fibers form intra-hemispheric connections across different lobes. Commissural tracts also establish connections between different cortical areas, but these specifically refer to inter-hemispheric connections. The projection tracts connect the cortex to other parts of the brain, such as deep nuclei, brain stem, cerebellum, and spinal cord. The tracts that interconnect the gray matter structures of the limbic system could be mainly categorized as projection or association tracts under this system of classification.

Advanced MRI techniques, particularly diffusion MRI, have afforded significant insights into the anatomy of the limbic system in the human brain, by enabling non-invasive and threedimensional mapping of structural connectivity *in vivo*. Both gray and white matter components of the limbic system have been studied by using relaxation-based (T1- or T2-weighted) MRI in Alzheimer's disease (Smith et al., 1999; Callen et al., 2001) as well as other neurological disorders affecting the limbic system (Atlas et al., 1986; Ng et al., 1997; Kuzniecky et al., 1999; Oikawa et al., 2001; Tsivilis et al., 2008; Lövblad et al., 2014). Diffusion MRI, which is based on sensitization to the directional-dependence of NMR signal attenuation arising from restricted diffusion of water molecules in brain tissue (Moseley et al., 1990; Le Bihan, 2003), has been used to investigate the structural connectivity of white matter tracts in the limbic system (Yamada et al., 1998; Wakana et al., 2004; Concha et al., 2005; Kalus et al., 2006; Malykhin et al., 2008; Zeineh et al., 2012), as well as to examine limbic pathways under pathologic conditions (Haznedar et al., 2000; Hattingen et al., 2007; Dineen et al., 2009; Wilde et al., 2010). Diffusion MRI employs a pair of diffusion-weighting gradients in the MR pulse sequence to impart sensitization to water diffusion in the brain which is influenced by surrounding tissue microstructure. Diffusion tensor imaging (DTI) uses a Gaussian

approximation of diffusion and is based on fitting the resulting signal decay to a six-element tensor model to estimate the degree and orientation of diffusion anisotropy, which can provide *in vivo* estimates and specific quantitative measures of white matter fiber structure and orientation (Basser and Jones, 2002; Tournier et al., 2011).

The limbic tracts in the human brain are relatively small in comparison to other mammalian species. For instance, major limbic tracts that can be identified with DTI in a mouse brain are demonstrated in **Figure 1**. It can be seen that these limbic tracts are large and well-defined, with the exception of the cingulum, which is relatively diffuse and characterized by an ambiguously defined boundary. As will be examined in detail in the following sections, many of these tracts are comparatively difficult to identify in the human brain. Interestingly, one exception is the cingulum, which is one of the most readily identifiable tracts in the human white matter (Burgel et al., 2006). **Figure 2** illustrates a connectivity diagram of the three major limbic tracts that are identifiable, albeit partially, in the human brain; namely, the cingulum, the stria terminalis, and the fornix. The Cg, which constitutes a part of the limbic system, receives sensory inputs from the neocortex (frontal, parietal, occipital, and temporal lobes) and projects to the hippocampal complex via the cingulum bundle. Anatomically, the cingulum forms a large outer C-shaped loop. The fornix also has a C-shaped trajectory that is nested within the cingulum, which is known to contain bidirectional connections between the septal area and the hippocampus. The stria terminalis forms the inner-most C-shaped trajectory connecting the septal area and the amygdala. The threedimensional reconstruction of these tracts based on deterministic fiber tractography from a human DTI study is shown in

**Figure 3**. The cingulum can be delineated coursing along the ventral surface of the hippocampal formation, while the fornix and stria terminalis project primarily along its dorsal surface (**Figure 3**).

# **MR IMAGING OF LIMBIC TRACTS IN THE HUMAN BRAIN**

Relaxometry-based and diffusion-weighted MR acquisitions generate complementary tissue contrasts for examination of limbic white matter anatomy, as will be shown in the following sections. T1-weighted imaging generally offers higher spatial resolution due to its relatively high signal-to-noise efficiency within clinically viable scan times. This can be specifically advantageous for visualizing the detailed morphology of the often highly convoluted limbic structures, such as the Cg and the hippocampal formation. One drawback of relaxation-based MRI is the lack of

high anatomical contrast to decipher the architecture of white matter structures and tracts in the limbic system. On the other hand, DTI can provide rich tissue contrast for visualizing white matter axonal architecture, based on the orientation of structural barriers (e.g., axonal membranes and myelin sheaths) that restrict water diffusion preferentially along directions orthogonal to the long axis of the axons. DTI, however, is limited in terms of the achievable spatial resolution *in vivo*. The inherent limitation on the spatial resolution for DTI stems from its high sensitivity to physiological motion and the ensuing necessity to use single-shot rapid imaging—constraints which can be circumvented to some extent for *ex vivo* DTI studies, thereby allowing higher spatial resolution acquisitions as shown in **Figure 1**. In the following sections, the anatomy of major limbic tracts delineated with MR studies of the human brain *in vivo* will be described in detail.

# **CINGULUM BUNDLE**

**Figure 4** shows a series of coronal sections from high-resolution T1-weighted imaging and DTI of the human brain. The T1 weighted images shown are acquired on a 7T MR scanner, using a magnetization-prepared rapidly-acquired gradient echo (MPRAGE) sequence with whole brain coverage and a matrix size of 368 × 368 × 261, resulting in an isotropic spatial resolution of 0.625 mm. The DTI data are acquired at 3T using a single-shot echo planer imaging (EPI) sequence, with matrix size of 128 × 128 × 72 and isotropic spatial resolution of 1.8 mm. The *b*-value

**FIGURE 4 | Coronal T1-weighted images (right panel) and corresponding DEC maps derived from DTI (left panel) showing major limbic structures and tracts delineated with MRI contrasts**. Images **(A–F)** show six coronal sections from the anterior to the posterior direction. Red, green, and blue in the DEC maps represent the primary orientation of diffusion along the medial-lateral, anterior-posterior, and superior-inferior axes, respectively. High-magnification views of select regions from T1-weighted contrasts are shown at the right, indicating the gray and white matter limbic structures that can be identified using MRI. Structural abbreviations are; c: cingulum, Ca: caudate, Cg: cingulate gyrus, Ct: tail of the caudate, fi: fimbria, fx: fornix, LGN: lateral geniculate nucleus, m: mammillothalamic tract, M: mammillary body, s: stria terminalis.

for DTI was 1000 s/mm<sup>2</sup> , with diffusion encoding applied along 32 non-collinear gradient directions.

The cingulum constitutes a compact bundle of both long and short association fibers that connect the cingulate cortex to the parahippocampal gyrus, prefrontal cortex, and cortical association areas in the parietal and occipital lobes (Schmahmann et al., 2007; Nieuwenhuys et al., 2008; Nezamzadeh et al., 2010). Delineation of the cingulum bundle based on MR contrasts across coronal sections from anterior to posterior (**Figures 4A–F**) can be seen in **Figure 4**. The cingulum can be readily identified in the direction-encoded color (DEC) maps in **Figures 4A–D**, marked by a distinct anterior-posterior orientation (green in DEC maps) adjacent to the cortical gray matter in the Cg and the corpus callosum. It makes an almost 180◦ U-turn ventrally around the splenium of the corpus callosum toward the temporal lobe, and as it turns to the superior-inferior direction, its color changes to blue in DEC contrasts (**Figure 4F**). After curving around the splenium, it projects along the inferior surface of the hippocampus and becomes smaller and more diffuse towards the anterior pole of the hippocampus (**Figures 4A–D**). The projection in the temporal lobe can also be appreciated in DEC maps in **Figures 4B–D**. The location of the cingulum can be estimated from corresponding T1-weighted images in coronal views (**Figures 4A–F**), however its boundary is less obvious and cannot be clearly distinguished from neighboring white matter tracts in T1-weighted contrasts.

#### **FORNIX**

The fornix is a projection tract that constitutes the major efferent fiber pathway from the hippocampal region, connecting it with the mammillary body, and then to the anterior thalamic nuclei through the mammillothalamic tract (Nolte, 1998; Nieuwenhuys et al., 2008; Thomas et al., 2011). This part of the Papez circuit, as well as the participating gray matter structures, can be precisely reconstructed from the high-resolution T1-weighted MR images as shown in **Figure 5**. The anatomical locations of these structures can also be appreciated in the coronal sections of T1-weighted

**body (via the fornix) and to the anterior thalamic nuclei (via the mammillothalamic tract) reconstructed from in vivo T1-weighted MRI of the human brain**.

images shown in **Figures 4A–F**. Fibers in the fornix arise from the hippocampus in each hemisphere, continue into the fimbria (delineated in T1-weighted contrasts in **Figures 4C,D**) and form the crus of the ipsilateral fornix. The crura continue forward and converge under the splenium of the corpus callosum to form the body of the fornix. Fimbrial fibers that continue medially across the midline to the contralateral hemisphere form the commissural component of the fornix known as the hippocampal commissure, which projects to the contralateral hippocampus, and is relatively less distinct in MR contrasts in the human brain in contrast to its large size and prominent delineation in mouse brains (**Figure 1**). Because of the relatively small size of this fine white matter tract, DTI contrasts can reveal only parts of the fornix in the brain *in vivo* (DEC maps in **Figure 4**). Several studies have examined pathological changes in the fornix associated with Alzheimer's disease (Oishi et al., 2012; Fletcher et al., 2013, 2014) and related disorders (Ng et al., 1997; Kuzniecky et al., 1999) using MRI. While the body and crus of the fornix are more apparent in T1-weighted and DEC contrasts, the anterior portion of the fornix as it approaches the septal and hypothalamic regions becomes more diffuse and narrow. Although there are several different anatomical targets of fibers in the fornix in this region, only the projection towards the mammillary body can be clearly delineated in high-quality T1-weighted images (**Figure 4**).

#### **STRIA TERMINALIS/FIMBRIA**

The stria terminalis is a limbic pathway that constitutes the major efferent connection from the amygdala to the septal nuclei and the hypothalamus (Parent, 1996). Because of its small size, the stria terminalis is relatively difficult to recognize in both T1-weighted and DTI contrasts, but can be identified sporadically at several locations. Along the majority of its trajectory, it travels adjacent to the medial surface of the caudate nucleus, all the way to the tip of the tail of the caudate as can be clearly seen in DTI contrasts in **Figures 4C–E**. Here again, T1 weighted images provide an estimate of the location of the stria terminalis in coronal sections, but its precise boundary cannot be demarcated due to the lack of clear tissue contrast within the white matter in T1-weighted MRI. The ventral portion of the stria terminalis is relatively difficult to observe with MRI. Interestingly, as the stria terminalis travels dorsally, along the roof of the inferior horn of the lateral ventricles, there is a region where the DTI contrast intensifies with a marked increase in fractional anisotropy (**Figure 4D**). Because this region is adjacent to the lateral geniculate nucleus (LGN), it is possible that fibers in this region are mixed with adjacent fibers from the optic radiation. The fimbria travels primarily along the dorsal surface of the hippocampus, which can be clearly identified in the high-resolution T1-weighted images (**Figures 4B–D**). Although the stria terminalis and the fimbria are anatomically separated along opposite banks of the inferior horn of the ventricles, it is difficult to distinguish them in this region with DTI, owing to limited spatial resolution and resulting partial volume effects in *in vivo* diffusion-weighted images. The stria terminalis continues to travel anteriorly, finally reaching the amygdaloid complex (**Figures 4A,B**). Although small, this section of the stria terminalis, not contaminated by adjacent fibers from the fimbria and optic radiation, can be clearly identified in high-quality DTI data (**Figures 4A,B**).

#### **CONCLUSION AND FUTURE DIRECTIONS**

In this mini review, we have described major tracts of the limbic system that can be delineated based on *in vivo* MRI of the human brain at clinical gradient strengths. The intricate threedimensional morphologies of gray and white matter structures and interconnecting pathways of the limbic circuitry that can be resolved using non-invasive MR methods are important for clinical studies and for evaluation of various neurologic disorders that affect the limbic system. Functional and structural neuroimaging studies continue to refine existing conception of the limbic system and its disorders. Advances in both MR hardware, including higher field strengths (7T–9.4T) and high-performance gradient systems (e.g., 300 mT/m) for human scanners, as well as diffusion MR acquisition and modeling techniques, will be instrumental in pushing the envelope of the achievable resolution and the level of structural detail for *in vivo* studies of the human brain, which can potentially afford additional insights into deciphering the complex circuitry of the limbic system.

#### **REFERENCES**


disorders studied with positron emission tomography and magnetic resonance imaging. *Am. J. Psychiatry* 157, 1994–2001. doi: 10.1176/appi.ajp.157.12. 1994


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 08 August 2014; accepted: 04 November 2014; published online: 27 November 2014*.

*Citation: Mori S and Aggarwal M (2014) In vivo magnetic resonance imaging of the human limbic white matter. Front. Aging Neurosci. 6:321. doi: 10.3389/fnagi.2014.00321*

*This article was submitted to the journal Frontiers in Aging Neuroscience*.

*Copyright © 2014 Mori and Aggarwal. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution and reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms*.

# Microstructure, length, and connection of limbic tracts in normal human brain development

#### **QiaowenYu1,2† ,Yun Peng<sup>3</sup>† , Virendra Mishra<sup>2</sup> , Austin Ouyang<sup>2</sup> , Hang Li <sup>3</sup> , Hong Zhang<sup>3</sup> , Min Chen<sup>4</sup> , Shuwei Liu<sup>1</sup>\* and Hao Huang2,5\***

<sup>1</sup> Shandong Provincial Key Laboratory of Mental Disorders, Research Center for Sectional and Imaging Anatomy, Shandong University School of Medicine, Jinan, China

<sup>2</sup> Advanced Imaging Research Center, University of Texas Southwestern Medical Center, Dallas, TX, USA

<sup>3</sup> Department of Radiology, Beijing Children's Hospital Affiliated to Capital Medical University, Beijing, China

<sup>4</sup> Department of Mathematical Sciences, University of Texas at Dallas, Richardson, TX, USA

<sup>5</sup> Department of Radiology, University of Texas Southwestern Medical Center, Dallas, TX, USA

#### **Edited by:**

Kenichi Oishi, Johns Hopkins University, USA

#### **Reviewed by:**

Daniel Ortuño-Sahagun, Centro Universitario de Ciencias de la Salud, Mexico Guo-Yuan Yang, Shanghai Jiao Tong University, China

#### **\*Correspondence:**

Shuwei Liu, Shandong Provincial Key Laboratory of Mental Disorders, Research Center for Sectional and Imaging Anatomy, Shandong University School of Medicine, No. 44 West Wenhua Road, Jinan, Shandong 250012, China e-mail: liusw@sdu.edu.cn; Hao Huang, Advanced Imaging Research Center, University of Texas Southwestern Medical Center, 5323 Harry Hines Blvd, Dallas, TX 75390-8542, USA e-mail: hao.huang@

utsouthwestern.edu

†Qiaowen Yu and Yun Peng have contributed equally to this work. The cingulum and fornix play an important role in memory, attention, spatial orientation, and feeling functions. Both microstructure and length of these limbic tracts can be affected by mental disorders such as Alzheimer's disease, depression, autism, anxiety, and schizophrenia. To date, there has been little systematic characterization of their microstructure, length, and functional connectivity in normally developing brains. In this study, diffusion tensor imaging (DTI) and resting state functional MRI (rs-fMRI) data from 65 normally developing right-handed subjects from birth to young adulthood was acquired. After cingulate gyrus part of the cingulum (cgc), hippocampal part of the cingulum (cgh) and fornix (fx) were traced with DTI tractography, absolute and normalized tract lengths and DTI-derived metrics including fractional anisotropy, mean, axial, and radial diffusivity were measured for traced limbic tracts. Free water elimination (FWE) algorithm was adopted to improve accuracy of the measurements of DTI-derived metrics. The role of these limbic tracts in the functional network at birth and adulthood was explored.We found a logarithmic age-dependent trajectory for FWE-corrected DTI metric changes with fast increase of microstructural integrity from birth to 2 years old followed by a slow increase to 25 years old. Normalized tract length of cgc increases with age, while no significant relationship with age was found for normalized tract lengths of cgh and fx. Stronger microstructural integrity on the left side compared to that of the right side was found. With integrated DTI and rs-fMRI, the key connectional role of cgc and cgh in the default mode network was confirmed as early as birth. Systematic characterization of length and DTI metrics after FWE correction of limbic tracts offers insight into their morphological and microstructural developmental trajectories. These trajectories may serve as a normal reference for pediatric patients with mental disorders.

**Keywords: limbic tract, development, trajectory, length, microstructure, DTI, connectivity, free water elimination**

#### **INTRODUCTION**

The limbic system is a group of interconnected cortical and subcortical structures dedicated to linking visceral states and emotion to cognition and behavior (Mesulam, 2000). The limbic system plays a key role in a variety of psychiatric and neuropsychological disorders including schizophrenia (e.g.,Torrey and Peterson, 1974; Reynolds, 1983; White et al., 2008; Leech and Sharp, 2014), major depressive disorder (e.g., Monkul et al., 2007; DeRubeis et al., 2008; Duman and Voleti, 2012), autism (e.g., Amaral et al., 2008), Alzheimer's disease (e.g., Braak and Braak, 1991), and obsessivecompulsive disorder (e.g., Fitzgerald et al., 2005;Abramowitz et al., 2009). As major white matter tracts underlying connectivity of the limbic system, cingulum bundle (including cingulum cingulate gyrus part or cgc and cingulum hippocampal part or cgh) and fornix (fx) could be affected in the pathological state. Moreover, structural changes of the limbic tracts could serve as an important

biomarker for early detection of these disorders as previous studies suggest that disruption of limbic tracts precedes the clinical symptoms. For example, disruption of cgc was found in healthy subjects with high risk of major depressive disorder (e.g., Huang et al., 2011). Microstructural changes of fx, a major fiber bundle connecting to hippocampus, was found in subjects of mild cognitive impairment (MCI), a disorder that has been associated with risk for dementia (e.g., Oishi et al., 2011, 2012; Huang et al., 2012). Besides their important role in aging-related diseases such as dementia, disruption of limbic tracts was also found to be associated with psychiatric disorders of developing brain, such as schizophrenia (e.g.,White et al., 2007) and autism (e.g.,Weinstein et al., 2011). Quantitative characterization of these limbic tracts systematically during development thus offers a normal reference for mental disorders. Furthermore, understanding maturation of the limbic tracts may also provide insights on vulnerability of these tracts in dementia of aging brains. Although the uncinate fasciculus and anterior thalamic projections could also be considered as part of the limbic system white matter tracts (e.g., Catani et al., 2013), cgc/cgh and fx were focused in this study.

In the developing human brain, water molecules tend to diffuse more freely along the limbic tract, instead of perpendicular to it. This diffusion property can be measured non-invasively with diffusion MRI, a modality of MRI. Diffusion tensor imaging (DTI) (Basser et al., 1994) characterizes water diffusion properties with a tensor model. DTI-derived metrics are highly sensitive to white matter microstructural changes. Fractional anisotropy (FA) (Pierpaoli and Basser, 1996; Beaulieu, 2002), derived from DTI and quantifying the shape of diffusion tensor, has been widely used to characterize microstructural integrity. Radial diffusivity (RD) and axial diffusivity (AD), also derived from diffusion tensor, convey unique information related to myelination and axonal integrity, respectively (Song et al., 2002). Mean diffusivity (MD) is the linear combination of AD and RD. Limbic tracts can be non-invasively traced with tractography based on diffusion MRI (e.g., Conturo et al., 1999; Jones et al., 1999a; Mori et al., 1999; Basser et al., 2000; Catani et al., 2002).

With traced limbic tracts as binary masks for DTI-derived metric maps, the microstructural properties of the limbic tracts can be quantified. Characterization of structural development of limbic tracts has been included in recent studies on white matter development (e.g., Hermoye et al., 2006; Lebel and Beaulieu, 2011). With different developmental period focused in these studies, different curves for trajectories of limbic tract microstructural changes were reported. Thus, far systematic microstructural measurements of limbic tracts with age coverage from birth to 25 years old has not been found in the literature. Moreover, corrections of bias of DTI metric measurements caused by partial volume effects (PVE) have not been reported either. As limbic tracts are relatively thin bundles compared to other major tracts such as corpus callosum and pathways of cgh and fx are close to the ventricle, contamination of cerebrospinal fluid (CSF) on DTI metric measurements could be severe. Such contamination may not be consistent during development as more water content was found in the younger brains such as neonate brain (e.g., Neil et al., 1998; Mukherjee et al., 2001, 2008). Corrections need to be made for a more accurate estimate of developmental trajectories of limbic tract DTI metrics. To the best of our knowledge, morphology of limbic tracts has not been quantitatively characterized. As coherent fiber bundles, it is possible to measure geodesic lengths of limbic tracts based on tractography results.

The limbic tracts are also involved in functional connectivity. It has been reported that cgc and cgh play a vital role in adult brain default mode network (DMN) (van den Heuvel et al., 2008, 2009; Greicius et al., 2009; Uddin et al., 2009). Specifically, cgc connects medial prefrontal cortex (MPFC) and posterior cingulate cortex (PCC) and cgh connects PCC and medial temporal lobes (MTL). Functional connectivity in DMN is related to episodic memory, theory of mind, and the ability to "mentalize" (Fair et al., 2008). Appearance of functional connectivity in DMN has been reported as early in development as the neonate stage (Doria et al., 2010). But weak relationship between functional and structural connectivity within the DMN exists during childhood (Supekar et al., 2010). To understand the relationship between structure and function, it is important to confirm involvement of limbic tracts in connecting functional regions in the DMN as early as birth.

In this study, DTI from 65 normally developing right-handed subjects at cross-sectional ages from birth to young adulthood and resting state functional MRI (rs-fMRI) from 15 of the subjects were acquired. After cgc, cgh, and fx were traced with DTI tractography, absolute and normalized tract lengths and DTI-derived metrics including FA, MD, AD, and RD were measured for traced limbic tracts. Free water elimination (FWE) algorithm (Pasternak et al., 2009) was adopted to enhance accuracy of the measurements of DTI-derived metrics. The role of these limbic tracts in the functional network at birth and adulthood was explored. Systematic characterization of tract geodesic length and DTI metrics after FWE correction of limbic tracts offers insight into their morphological and microstructural developmental trajectories. These trajectories may serve as a normal reference for pediatric patients with mental disorders.

## **MATERIALS AND METHODS HUMAN SUBJECTS**

Sixty-five healthy human subjects (41 M/24 F) with age range from 0 month (birth) to 25 years participated in this study. The age range covered the important developmental periods of early childhood, childhood, adolescence, and young adulthood. **Figure 1** shows a histogram of the age and gender distribution of all subjects. Included neonates were part of the cohort for studying normal perinatal and prenatal brain development and were selected after rigorous screening procedures (Huang et al., 2014). Neonates were well fed before scanning. All included children, adolescents and young adults were healthy subjects free of current and past neurological or psychiatric disorders. Right-handed were reported for all children who showed clear handedness. For young children, besides earplugs and earphones, extra foam padding was applied to reduce the sound of the scanner while they were asleep. Due to difficulty of acquiring MRI from human subjects with such a comprehensive developmental period from 0 month to

25 years in one site, the presented data were acquired from the same type of MR scanner with common MRI protocol in two sites, Children's Medical Center at Dallas and Children's Hospital in Beijing. All subjects gave informed written consent approved by the Institutional Review Board of those two institutions.

#### **DATA ACQUISITION OF DTI AND fMRI**

MRI of 3 neonates (2 M/1 F) at 40 weeks of gestation and 10 adolescents (12–17 years old; 4 M/6 F) was acquired from Children's Medical Center at Dallas. MRI of all other subjects was acquired from Children's Hospital in Beijing. A Philips 3 T Achieva MR scanner at both sites was used. A single-shot echo planar imaging (EPI) with SENSE (SENSitivity Encoding, reduction factor = 2.5) parallel imaging scheme was adopted for DTI acquisition. Eight-channel SENSE head coil and consoles installed with R2.6.3 software were used for both sites. The DTI imaging parameters were: TR (repetition time) = 6850 ms, TE (echo time) = 78 ms, in-plane field of view = 168 mm × 168 mm for neonates and 224 mm × 224 mm for all others, in-plane imaging resolution = 1.5 mm × 1.5 mm for neonates and 2 mm × 2 mm for all others, slice thickness = 1.5 mm for neonate to 2 mm for all others, slice number = 60–65 depending on the height of the subject brains, 30 independent diffusion-weighted directions (Jones et al., 1999b) uniformly distributed in space, *b*value = 1000 s/mm<sup>2</sup> , and repetition = 2. The axial image dimension was 256 × 256 after reconstruction. The total acquisition time of DTI was 11 min. With 30 diffusion weighted image (DWI) volumes and 2 repetitions, we accepted only those scanned datasets with less than 5 DWI volumes affected by motion more commonly seen in scan of neonates and young children. The affected volumes were replaced by the good volumes of another DTI repetition during postprocessing. rs-fMRI was also acquired from 15 out of 65 subjects, including 3 neonates and 3 young adults. A T2\* weighted gradient echo EPI sequence was used. A total of 210 T2\* weighted whole brain volumes were acquired with the following parameters: TR = 1500 ms for neonates and 2000 ms for young adults, TE = 27 ms for neonates and 30 ms for young adults, flip angle = 80°, in-plane imaging resolution = 2.4 mm × 2.4 mm for neonates and 3.4 mm × 3.4 mm for young adults, and slice thickness = 3mm with no gap for neonates and 4 mm with no gap for young adults; slice number = 30 for neonates and 37 for young adults. Scan time of rs-fMRI was 5–6 min. Rs-fMRI and DTI images were acquired in the same session.

## **TENSOR FITTING AND DTI-BASED TRACTOGRAPHY**

Diffusion weighted images for each subject were corrected for motion and eddy current by registering all DWIs to the b0 image using a 12-parameter (affine) linear registration with automated image registration (AIR) algorithm (Woods et al., 1998). After the registration, six independent elements of the 3 × 3 diffusion tensor (Basser et al., 1994) were determined by multivariate leastsquare fitting of DWIs. The tensor was diagonalized to obtain three eigenvalues (λ1−3) and eigenvectors (*v*1−3). FA, MD, AD, and RD, derived from DTI and used to quantify white matter microstructure, were obtained for all the subjects using the equations of eigenvalues described in details previously (e.g., Mishra et al., 2013).

Fiber assignment of continuous tractography (FACT) (Mori et al., 1999) was used to trace limbic tracts for all subjects. The limbic tracts include cgc, cgh, and fx. FA threshold of 0.2 and a principal eigenvector turning angle threshold of 50 were used for FACT tractography. cgc and cgh were traced following the fiber tracking protocol (Wakana et al., 2007) with a multiple-ROI approach (Huang et al., 2004). The ROI placement protocol for tracing fx was as follows. The first ROI was drawn to include main body of fx identifiable on an axial slice of DTI color-encoded map. The second ROI was placed on the coronal slice where hippocampus was identifiable using "AND" operation. All other fibers were removed carefully to retain only the fx. The above-mentioned tensor fitting and fiber tracking was conducted by using DTIStudio (Jiang et al., 2006).

#### **MEASUREMENTS OF LENGTHS AND DTI METRICS OF LIMBIC TRACTS**

After tractography of cgc, cgh, and fx, the lengths and DTI metrics of each tract were measured for all subjects and plotted against age. These limbic tracts were visualized in 3D with Amira software (FEI, Berlington, MA, USA).

#### **Measurement of tract length**

The absolute and normalized limbic tract length were measured based on the tractography results. Specifically, geodesic length of each fiber in cgc, cgh, and fx was measured. A histogram was established for each limbic tract to show the distribution of fiber length. Note that a tract contains fibers with different lengths. In order to reduce the bias caused by a large number of short fibers, we defined the tract length by averaging 10% longest fibers of each tract. As shorter fibers could only cover part of the tract, reconstruction of the top 10% longest fibers showed that morphology of these fibers was in good agreement with that of the entire tract of all fibers. Tract length was further normalized by the size of the brain represented by the length between most anterior to most posterior edge to quantify relative tract length change during development.

#### **Measurements of uncorrected and corrected DTI-based metrics**

The binary masks of the individually traced limbic system tracts were used to compute the tract-level FA, MD, AD, and RD without correction of PVE of free water, defined as uncorrected DTI metrics. Young brains contain a large amount of free water. The uncorrected DTI metrics reflect the weighted average of all compartments including free water. In order to minimize the influence of CSF, FWE (Pasternak et al., 2009) was applied to correct the DTI metrics of cgc, cgh, and fx by using the following model: *S* = *S*<sup>0</sup> *f*iso*e* <sup>−</sup>*bD*iso + (1 − *f*iso)*e* −*bD*¯ , where *S* is the measured signal in the voxel, *f*iso is the volume fraction of the isotropic compartment, *b* is the applied *b*-value, *S*<sup>0</sup> is the signal with no diffusion weighting, *D*iso is the diffusion coefficient of the isotropic compartment, and ¯*D*¯ is the corrected tensor representing the fiber bundle in the voxel. This is a two-compartment model with one compartment representing the fiber bundle and the other compartment representing the isotropic free water compartment in the voxel. We estimated the six elements of ¯*D*¯ and *<sup>f</sup>*iso by solving the non-linear equation in the model. The initial values of ¯*D*¯ for iteration are those from single tensor model without isotropic component. The diffusivity of the isotropic compartment *D*iso was 3 × 10−<sup>3</sup> mm<sup>2</sup> /s (Pasternak et al., 2009). Note that with measurements from 30 diffusion gradients, an over-determined system was used to fit the six independent elements of tensor ¯*D*¯ and *<sup>f</sup>*iso.

#### **Rs-fMRI DATA PROCESSING**

The DMNs of three neonates and three young adults were identified with independent component analysis (ICA) of the rsfMRI data. ICA was conducted with MELODIC tool (Multivariate Exploratory Linear Decomposition into Independent Components, a part of FSL)<sup>1</sup> . Pre-processing procedures included removing the first 10 volumes for stabilization of the magnetic field, motion correction, spatial smoothing with a Gaussian kernel of FWHM 5 mm, intensity normalization, and high-pass temporal filtering. Volumes with translational movement greater than 5 mm were removed from the dataset. The DMN cluster components in the fMRI space were transformed into the DTI space by registering the resampled fMRI image to b0 image in the DTI space using SPM8 (Statistical Parametric Mapping 8)<sup>2</sup> . Amira software (FEI, Berlington, MA) was used for reconstructing the DMN clusters and limbic tracts connecting them.

#### **CURVE FITTING AND STATISTICAL ANALYSIS**

The following equation was used for fitting a model between measurement *y* (including uncorrected/corrected DTI metrics and absolute/normalized tract length) and age *t*, *y* = *f*(*t*) + *c*·sex + ε, where *f*(*t*) was linear, logarithmic, or polynomial function of *t* and ε was an error term. *F*-test was used to find the best fitting curve. We first tested if the measurement *y* was age dependent, if so, we

further tested which of the following curves, linear, logarithmic, or polynomial curves, fitted the data best.

With age and gender as covariates, generalized linear model (GLM) was used to test (1) if FWE correction significantly changed DTI metric measurements and (2) if lateralization (i.e., measurement in left or right hemisphere) was a significant factor for all measurements. The null hypothesis for FWE correction is that there is no difference of DTI metric measurement before and after the correction. The null hypothesis for lateralization is that there is no difference of length or DTI metric measurement of corresponding limbic tracts in the left and right hemisphere. The Bonferroni correction was used to control the spurious positives when rejecting the null hypotheses.

#### **RESULTS**

#### **3D RECONSTRUCTED LIMBIC TRACTS AT DIFFERENT DEVELOPMENTAL STAGES**

As shown in **Figure 2**, overall consistent 3D reconstructed limbic tracts, cgc, cgh, and fx, can be observed at five time points throughout the developmental period from 0 month to 25 years. cgc appears to be more extended to prefrontal regions and has more branches during development, while morphology of cgh and fx remains relatively stable.

#### **DEVELOPMENT OF MICROSTRUCTURES OF LIMBIC TRACTS**

The plots of uncorrected and corrected DTI metrics, namely, FA, MD, AD, and RD, of all limbic tracts at different ages are shown in **Figures 3**–**6**, respectively. As shown in **Tables 1**–**4**, significant age dependence was found in almost all DTI metrics of all limbic tracts, except measurements of MD, AD, and RD of fx. Among logarithmic, linear, and polynomial fitting between uncorrected/corrected DTI metrics and age, logarithmic model fits best to most of DTI

<sup>1</sup>www.fmrib.ox.ac.uk/fsl/melodic2/index.html <sup>2</sup>http://www.fil.ion.ucl.ac.uk/spm/

listed in each panel.

**FIGURE 4 | Age-dependent changes of uncorrected and corrected mean diffusivity (MD) measurements of limbic tracts**. R <sup>2</sup> and fitting equations are listed in each panel.

metric measurements of the limbic tract, also shown in **Tables 1**– **4**. For rest of the fittings, there is no statistical difference between linear and logarithmic fitting. No significant difference between linear and logarithmic fitting is especially apparent for corrected FA measurements of all limbic tracts (**Table 1**). Lowest *R* <sup>2</sup> was obtained with the polynomial fitting.

in each panel.

**FIGURE 6 | Age-dependent changes of uncorrected and corrected radial diffusivity (RD) measurements of limbic tracts**. R <sup>2</sup> and fitting equations are listed in each panel.

As shown in **Table 5**, DTI metrics of all limbic tracts were significantly altered after FWE correction. FA increases due to FWE correction for fx is apparent in **Figure 3**, while uncorrected and corrected FA of cgc-L/R and cgh-L/R overlap with no visible changes. Decrease of MD, AD, and RD of all limbic tracts can be clearly observed in **Figures 3**–**6**. Statistical results in **Table 5**

#### **Table 1 | Relationship between fractional anisotropy (FA) of limbic tracts and age**.


Statistical significance p < 0.05, p < 0.01, and p < 0.0001 are indicated by bold \*, \*\*, and \*\*\*, respectively. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; UC, uncorrected; C, corrected; L, left; R, right; NS, not significant.

#### **Table 2 | Relationship between mean diffusivity (MD) of limbic tracts and age**.


Statistical significance p < 0.05, p < 0.01, and p < 0.0001 are indicated by bold \*, \*\*, and \*\*\*, respectively. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; UC, uncorrected; C, corrected; L, left; R, right; NS, not significant.

indicate that there are statistically significant changes for all DTI metrics of all limbic tracts with FWE correction.

#### **DEVELOPMENT OF LENGTHS OF LIMBIC TRACTS**

The plots of absolute and normalized lengths of all limbic tracts at different ages are shown in **Figure 7**. As shown in **Table 6**, significant age dependence was found in absolute lengths of all limbic tracts and normalized lengths of cgc-L/R. Among logarithmic, linear, and polynomial fitting between measured length and age, polynomial model was first rejected due to lowest *R* 2 . After *F*-test,

#### **Table 3 | Relationship between axial diffusivity (AD) of limbic tracts and age**.


Statistical significance p < 0.05, p < 0.01, and p < 0.0001 are indicated by bold \*, \*\*, and \*\*\*, respectively. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; UC, uncorrected; C, corrected; L, left; R, right; NS, not significant.



Statistical significance p < 0.05, p < 0.01, and p < 0.0001 are indicated by bold \*, \*\*, and \*\*\*, respectively. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; UC, uncorrected; C, corrected; L, left; R, right; NS, not significant.

significant difference between logarithmic and linear model was found only for absolute length of cgc-L (**Table 6**). Logarithmic curves were fitted with highest *R* 2 for absolute and normalized lengths of all limbic tracts, as shown in dashed and solid curves in **Figure 7**, respectively. **Figure 7** and **Table 6** suggest that increases of absolute tract lengths of cgh-L, cgh-R, and fx results from overall brain size increases during development, while increases of absolute tract length of cgc-L and cgc-R are due to both overall brain size increases and relative increases of lengths of these two tracts in the brain.


**Table 5 | Comparisons of DTI metrics of limbic tract before and after FWE correction with age and gender as covariates**.

Statistical significance after Bonferroni correction is indicated by bold. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; L, left; R, right; \*\*\*p < 0.001.

#### **LATERALIZATION**

Higher FA and lower AD, RD, and MD in the left cgc/cgh can be observed in **Figures 3**–**6**. In addition, longer normalized tract length of cgc on the left side compared to that on the right side can be observed in **Figure 7**. Statistical comparisons of normalized tract length and corrected DTI metrics on the left and right side are listed in **Table 7**. Significant differences were found in all DTI metrics between the left and right cgc and between the left and right cgh (**Table 7**). Statistically significant difference of normalized tract length was found between cgc-L and cgc-R, but not between cgh-L and cgh-R (**Table 7**).

#### **STRUCTURAL CONNECTIVITY OF DMN REGIONS THROUGH LIMBIC TRACTS FOR NEONATE AND ADULT BRAIN**

**Figure 8** shows the functional connectivity maps in the DMN and structural limbic tracts connecting the DMN regions for a neonate and an adult brain. Consistency of DMN regions (MPFC, PCC, and MTL) and limbic tracts connecting these regions is clear between the neonate (**Figure 8A**) and adult (**Figure 8B**) brain.

Specifically, cgc connects MPFC and PCC; cgh connects PCC and MTL for both neonate and adult brain.

# **DISCUSSION**

With DTI from 65 subjects of cross-sectional age from 0 month to 25 years and rs-fMRI from 15 of them, we have quantitatively characterized development of microstructure, length, and connection of limbic tracts that play a key role in emotion, memories, and behavior. Limbic tracts were traced with DTI tractography. As limbic tracts are relatively thin and discrete compared to other major white matter tracts such as corpus callosum, the contamination of CSF to DTI metric measurements cannot be ignored. Correction of DTI metric measurement with FWE was conducted to achieve more accurate measurements of limbic tract microstructure. Together with the tract length measurement, microstructure and length of limbic tracts from birth to adulthood were quantitatively characterized. In addition, consistent role of limbic tracts connecting major DMN regions at birth and adulthood has been confirmed and demonstrated. Comparisons between DTI metrics and lengths of the left and right limbic tracts suggest significant lateralization with longer and better myelinated cgc on the left side compared to that on the right side.

Despite that the overall shape of limbic tracts is relatively consistent throughout the developmental period from birth to young adulthood (**Figure 2**), **Figures 3**–**6** demonstrate rapid increase of FA and decrease of AD, RD, and MD within the first 2 years and relatively slow changes of these DTI metrics after 2 years until a plateau is reached during adolescence. Age-dependent increase of FA could be caused by larger decrease of RD compared to that of AD as FA is roughly the ratio between AD and RD. Rapid FA changes of cingulum within first 2 years could be caused by myelination, the majority of which begins within the first 2 years of life (Yakovlev and Andre-Roch, 1967). The asynchronous myelination of white matter tracts has been reported in neuropathology (Kinney et al., 1988) and MRI study (e.g., Deoni et al., 2011). Memory,

#### **Table 6 | Relationship between limbic tract length (absolute and normalized) and age**.


Statistical significance is indicated by bold. cgc, right cingulate gyrus part of cingulum; cgh, hippocampal part of cingulum; fx, fornix; L, left; R, right; \*p < 0.05; \*\*\*p < 0.001; NS, not significant.

emotion, and motivation functions are related to limbic tracts and important for survival. It is vital for limbic tracts to become well myelined earlier than other tracts, especially those projected from frontal and temporal lobes (Baumann and Pham-Dinh, 2001).

Including cross-sectional ages from birth to 25 years with a consistent group of dataset could offer more complete insight into structural maturation of limbic tracts. After test of three models including linear, logarithmic, and polynomial, **Tables 1**–**4** show that logarithmic model is overall best fitted to most limbic tract microstructural trajectories. Microstructural measurements of developing limbic tracts were found in previous literature (e.g., Hermoye et al., 2006; Grieve et al., 2011; Lebel and Beaulieu, 2011), which includes less comprehensive developmental periods. Depending on the early or late developing time window focused in these previous studies, different curves for developmental trajectories of limbic tracts were reported. For example, a non-linear trend of DTI metrics of white matter tracts was reported in a longitudinal DTI study with age range from 5 to 32 years (Lebel and Beaulieu, 2011). In that study, a large number of subjects were enrolled with 2 scans for most of the subjects, but early childhood period from 0 to 5 years was absent. During the early stage of development from birth to 54 months, non-linear trends could also be observed from measurements of FA and averaged apparent diffusion coefficient of cingulum (Hermoye et al., 2006), showing rapid DTI metric change in the first 3–6 months. A linear model was reported to be more suitable for very early development of limbic tracts of infants with shorter age range from 3.9 to 18.4 weeks (Dubois et al., 2008). In another study with later stage of limbic tract development from childhood to adulthood, polynomial curve was reported (Grieve et al., 2011). These previous findings are generally in agreement with the results presented in this study, since segmented developmental trajectory of DTI metrics at different developmental time window from birth to childhood could be quite different. For example, it can be observed from **Figures 3**– **6** that limbic tract metric changes before 2 years demonstrate clear linear behavior.

cgc-L/R, cgh-L/R, and fx are relatively thin and discrete white matter tracts. Contamination of brain CSF to DTI metric measurements of the limbic tracts could be severe. We adopted FWE (Pasternak et al., 2009) to achieve a more accurate characterization of developmental trajectories of microstructural measures. It



Statistical significance p < 0.001 after Bonferroni correction is indicated by bold \*\*\*. L, left; R, right; NS, not significant.

**coronal slices from the left to right show similar connectivity pattern in neonate (A) and adult (B) brain**. The color of connected brain clusters encodes t values. 3D visualization on the most right panels reveals clearly

cingulum hippocampal part (cgh) connects PCC and MTL for both neonate and adult brain. The 3D reconstruction of clusters at MPFC, PCC, and MTL was directly extracted from rs-fMRI analysis.

is clear that FA is increased and AD, RD, and MD are decreased after FWE correction, causing clear shift of DTI metric trajectories in **Figures 3**–**6**. Results in **Table 5** confirmed statistical significance of the shift after FWE correction. The amount of correction of DTI metric measurement is more dramatic for fx compared to that of cgc or cgh, as fibers of fx course closely around the ventricle and metric measurements of fx could be affected by severe CSF contamination. FA trajectories of cgc-L/R and cgh-L/R (**Figure 3**) are less affected by FWE correction than those of MD, AD, and RD (**Figures 3**–**6**) as FA can be considered approximately as a ratio of AD and RD and shift of AD and shift of RD offset each other. The shifts between uncorrected and corrected curves in **Figures 3**–**6** were heterogeneous. Specifically, DTI metric measurements of cgc-L/R and cgh-L/R were more severely affected by free water contamination in early childhood (**Figures 3**–**6**). Nevertheless, these shifts did not change the fitting models for most limbic tracts (**Tables 1**–**4**). Only for corrected FA of cgc-L/R and cgh-L/R, developmental trajectories of them could also be fitted with linear model as well as logarithmic model (**Table 1**) after FWE correction. To the best of our knowledge, no FWE correction has been reported for previous DTI metric measurements of developing limbic tracts. It is possible that FWE correction could help restore the real underlying microstructural maturation process of limbic tracts. Nevertheless, further cross-validation from a larger sample dataset is still needed to ensure accuracy of DTI metric measures after FWE correction.

Lengths of all limbic tracts increase during development, as shown in **Figure 7** and **Table 6**. Specifically, the length increases of cgh-L/R and fx follow the overall brain size development and no significant age dependence has been found after these lengths were normalized by the brain size. Significant age dependence has been found only for normalized lengths of cgc-L/R, indicating that there are extra length increases of cgc-L/R besides following growth of entire brain. Although the overall shape of cgc is relatively stable throughout development, extra cgc growth can be observed in its anterior part close to prefrontal cortex (**Figure 2**). Relative increase of cgc length is probably related to its growth in the prefrontal region. Functions of prefrontal areas are involved in planning, decision making, and moderating social behavior that develop during late childhood and adolescence (e.g., Gogtay et al., 2004). Connection of anterior cgc to late-developed prefrontal cortex may explain extra length growth of cgc from birth to young adulthood, but no extra length growth of cgh or fx.

Significant lateralization has been found for all DTI metrics of cgc-L/R and cgh-L/R with age and gender as covariates, as shown in **Table 7**. Higher FA and lower AD, RD, and MD of cgc and cgh can be observed on the left side than right side (**Figures 3**–**6**). This lateralization was associated with higher microstructural integrity on the left side of limbic tracts. Lateralization of DTI metrics of cgc and cgh may be related to unique functions of the left side of human brain such as language (van Veen et al., 2001). Exclusive right-handedness of the recruited subjects may also play a role. These findings are consistent to previous DTI metric measurements of cingulum (Gong et al., 2005; Verhoeven et al., 2010). Strengthened left limbic tracts were also demonstrated by longer length of cgc-L compared to that of cgc-R (**Table 7**). However, no significant length difference could be found for the left and right cgh.

In functional connectivity, brain regions where spontaneous blood oxygen level-dependent (BOLD) signal oscillations are temporally correlated can be either directly connected by white matter tracts or connected through a relay. The adult brain regions of PCC, MPFC, and MTL in DMN are connected by limbic tracts (Greicius et al., 2009). **Figure 8** clearly demonstrates consistent connectivity roles of limbic tracts in connecting DMN regions for both the neonate and adult brain. Immature DMN functional connectivity has been reported with rs-fMRI of neonate brains recently (Fransson et al., 2007, 2009; Doria et al., 2010; Smyser et al., 2010). Unlike other major white matter tracts such as arcuate fasciculus, limbic tracts are well formed at birth, as shown in **Figure 2**. However, the role of limbic tracts in connecting DMN regions as early as birth time has not been confirmed previously. With myelination of cgc and cgh during development indicated by **Figures 3**–**6**, it can be speculated that enhanced myelination of cgc and cgh provide structural basis for maturation of DMN during development.

It is difficult to acquire all DTI data from birth to young adulthood in one site. We combined datasets acquired in two sites, but with the same imaging protocol and same Philips 3 T Achieva scanners. Rigorous quality control is conducted routinely for both scanners. We performed quantitative analysis based on DTI tractography and metric measurements to evaluate the effects of these two scanners in our previous study (Huang et al., 2014), and it was concluded that the scanner effects are negligible. Multiple factors could contribute to variations of MRI data, including variability caused by thermal variations of the same scanner itself, physiological variability of the participant and differences of two scanners. In an earlier experiment, DTI was acquired from the same young adult human volunteer scanned at two Philips 3 T Achieva scanners. FA measurement differences caused by scanner difference were tested to be within the range of variability of scanning the same subject twice with one scanner (Saxena et al., 2012). These previous experiments suggest that scanner difference will not cause significant differences of reported results. It also should be noted that this study is not a real longitudinal one as the subjects at each cross-sectional age were not the same. It will be very difficult to follow a same cohort of subjects from birth to 25 years old. However, multiple scans from same subjects during their development could be helpful to characterize real longitudinal development of limbic tracts.

In conclusion, with DTI and rs-fMRI from developing brains at cross-sectional ages from 0 month to 25 years, we have quantitatively characterized development of microstructure, length, and connection of limbic tracts that play a key role in emotion, memories, and behavior. Logarithmic developmental trajectories were found for most age-dependent changes of limbic tract microstructures and lengths from birth to young adulthood. Correction of CSF contamination has significantly increased FA and reduced MD, AD, and RD of developing limbic tracts, but has not changed the age-dependent trajectories of these metrics. Stronger microstructural integrity and longer tract lengths were found for limbic tracts on the left side compared to those on the right side. Consistent role of limbic tracts connecting major DMN regions at birth and adulthood has been confirmed and demonstrated.

#### **ACKNOWLEDGMENTS**

This study is sponsored by NIH (MH092535 and MH092535- S1), Natural Science Foundation of China (Grant Nos. 31271161 and 31071050), and Specialized Research Fund for the Doctoral Program of Higher Education of China (No. 20120131130008).

#### **SUPPLEMENTARY MATERIAL**

The Supplementary Material for this article can be found online at http://www.frontiersin.org/Journal/10.3389/fnagi.2014.00228/ abstract

#### **REFERENCES**


diffusion anisotropy measured by using diffusion tensor MR imaging. *Radiology* 209, 57–66. doi:10.1148/radiology.209.1.9769812


intramodality validation. *J. Comput. Assist. Tomogr.* 22, 139–152. doi:10.1097/ 00004728-199801000-00028

Yakovlev, P. I., and Andre-Roch, L. (1967). "The myelogenetic cycles of regional maturation of the brain," in *Regional Development of the Brain in Early Life*, ed. A. Minkowski (Oxford: Blackwell Scientific), 3–70.

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 10 May 2014; accepted: 08 August 2014; published online: 28 August 2014.*

*Citation: Yu Q, Peng Y, Mishra V, Ouyang A, Li H, Zhang H, Chen M, Liu S and Huang H (2014) Microstructure, length, and connection of limbic tracts in normal human brain development. Front. Aging Neurosci. 6:228. doi: 10.3389/fnagi.2014.00228*

*This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2014 Yu, Peng , Mishra, Ouyang , Li, Zhang , Chen, Liu and Huang . This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Fornix as an imaging marker for episodic memory deficits in healthy aging and in various neurological disorders

# *Vanessa Douet\* and Linda Chang\**

*Department of Medicine, John A. Burns School of Medicine, University of Hawaii, Honolulu, HI, USA*

#### *Edited by:*

*Jean Mariani, Universite Pierre et Marie Curie, France*

#### *Reviewed by:*

*Olivier Piguet, Neuroscience Research Australia, Australia Claudia Metzler-Baddeley, Cardiff University, UK*

#### *\*Correspondence:*

*Vanessa Douet and Linda Chang, Department of Medicine, John A. Burns School of Medicine, The Queen's Medical Center, 1356 Lusitana Street, UH Tower, Room 716, Honolulu, HI 96813, USA e-mail: douet@hawaii.edu; lchang@hawaii.edu*

The fornix is a part of the limbic system and constitutes the major efferent and afferent white matter tracts from the hippocampi. The underdevelopment of or injuries to the fornix are strongly associated with memory deficits. Its role in memory impairments was suggested long ago with cases of surgical forniceal transections. However, recent advances in brain imaging techniques, such as diffusion tensor imaging, have revealed that macrostructural and microstructural abnormalities of the fornix correlated highly with declarative and episodic memory performance. This structure appears to provide a robust and early imaging predictor for memory deficits not only in neurodegenerative and neuroinflammatory diseases, such as Alzheimer's disease and multiple sclerosis, but also in schizophrenia and psychiatric disorders, and during neurodevelopment and "typical" aging. The objective of the manuscript is to present a systematic review regarding published brain imaging research on the fornix, including the development of its tracts, its role in various neurological diseases, and its relationship to neurocognitive performance in human studies.

**Keywords: fornix, development, aging, episodic memory, neuropsychiatric disorders, DTI**

# **INTRODUCTION**

The fornix is part of the limbic system that comprises cortical and subcortical structures. The cortical structures include cingulate and parahippocampal gyri, as well as the entorhinal cortex. The subcortical structures comprise the amygdalae, septal nuclei, nucleus accumbens, mammillary bodies, hypothalamus, anterior nucleus of the thalami, hippocampi and fornix.

The limbic system was first described by Pierre Paul Broca [1827–1880] (Broca, 1890) and was proposed to be the circuit of emotional experience and behavior by James W. Papez [1883–1958] (Papez, 1937). Later, its functions were linked to pleasure and reward, as well as memory and integration of memories (Rajmohan and Mohandas, 2007). Episodic memory belongs to the long-term memory system, and refers to conscious recollection of specific events (episodes) and contexts (time and place). Episodic memory frequently declines with aging and often becomes deficient in neurodegenerative diseases (Samson and Barnes, 2013) and psychiatric disorders (White et al., 2008). The critical subcortical structure for memory functions is the hippocampus (Penfield and Milner, 1958). As the major efferent white matter tract from the hippocampus, the fornix was frequently evaluated in relation to hippocampi and to memory impairments, especially to deficits in episodic memory (Yanike and Ferrera, 2014). Cumulative data from structural and diffusion tensor imaging (DTI) studies suggest that forniceal measures correlate with episodic memory performance in various neuropathological conditions, as well as during "typical" brain development and brain aging. The fornix appears to be a robust imaging predictor of episodic memory performance, independent of age and the etiology that may affect the integrity of the fornix.

In this review, we will focus on the findings from imaging studies of the fornix including its development, its implication in cognitive performance, and the structural changes associated with typical aging and neurodegenerative disorders. After a brief description of its anatomy, we will summarize the studies conducted on the forniceal formation across the lifespan, particularly those assessed by DTI which provided much new knowledge in our understanding of the fornix. We will then concentrate on diseases that may lead to an impaired or underdeveloped fornix and its likely consequences on cognitive performance, particularly in episodic memory.

#### **MATERIALS AND METHODS**

We searched in the PubMed® database for relevant publications during the last decade (last update on 2014 November 15th). Our search terms included "MRI," "DTI," diseases of interest, "aging," "development," "cognition," "memory" in combination with "fornix." 482 results were obtained. We screened the abstracts and included only the papers that were original, published in English and referred to human research. However, the most important selection criterion was that the studies explicitly reported imaging findings of the fornix. Conference abstracts and case reports were excluded. After screening all relevant studies and excluding those papers not fulfilling the inclusion criteria, we evaluated 143 studies in further detail.

#### **ANATOMY OF THE FORNIX (FIGURE 1)**

The fornix of the brain is a C-shaped structure that projects from the posterior hippocampus to the septal area and hypothalamus. As the hippocampus terminates near the splenium of the corpus callosum, the fimbria becomes a detached bundle, the crus of the fornix. The two crura merge medially to form the body of the fornix. At the interventricular foramen, the body of the fornix diverges into the two adjacent columns that pass through the middle of the hypothalamus toward the mammillary bodies (**Figure 1A**).

The fornix is the largest efferent pathway from the hippocampus, and belongs to the "Papez circuit," which is also referred to as the limbic system. Forniceal fibers from the forebrain project to the anterior nucleus of the thalami, the mammillary bodies, hypothalamus, the septal nuclei and the ventral striata. Some fibers of the precommissural fornix spread beyond the septal nuclei and the ventral striata, and reach the orbital and anterior cingulate cortices. Forniceal fibers also contact the entorhinal cortex, amygdalae and back-project to the posterior cingulate gyrus (Nolte, 2009). The Papez circuit, or the limbic system, is involved in learning, memory, emotion and social behavior (King et al., 2013).

#### **NORMAL DEVELOPMENT AND AGING OF THE FORNIX (TABLE 1)**

On T1-weighted MRI, the left and right columns of the fornix are difficult to delineate, and are mainly treated as a single central structure that diverges into both cerebral hemispheres. Forniceal changes are often associated with abnormalities in surrounding structures, resulting in structural distortions that are difficult to assess. However, DTI can differentiate more easily the fornix from surrounding structures, and can quantify microstructural changes within the fornix. DTI characterizes the three-dimensional diffusion of water molecules and provides information on the integrity of tissue microstructures. The fractional anisotropy (FA) value indicates the architectural degree of the tissue, which may be influenced by the amount of myelination, the coherence of axonal fibers, or a combination of both, while the mean diffusivity (MD) value is a measure of the overall averaged water diffusion within a volume of tissue. For instance, a lower than typical FA observed during development of healthy children might indicate hypomyelination or slower growth of the axons, while a decline in FA might reflect either demyelination or a decline in the number of myelinated axonal fibers, or both. An increase in MD is associated with either neuronal damage or degeneration of microstructural barriers such as cell membranes. Loss of myelin typically increases radial diffusivity (RD), whereas axial diffusivity (AxD) may be a more specific marker of axonal damage (Song et al., 2002, 2003).

Recently, the development of the fornix across the lifespan has become an active area of investigation because of the quality of visualization which is possible with DTI. However, only a small minority of these studies was conducted longitudinally (**Table 1**).

On the post-mortem human fetal brain, the fornix can be identified on MRI as early as 10 weeks of gestation (Rados et al., 2006). DTI techniques showed that the fornix is one of the most prominent tracts in the fetal brain and its entire tract is fully formed by 13 gestational weeks (Huang et al., 2006, 2009).

#### **Table 1 | Fornix metrics across the lifespan.**


#### **Table 1 | Continued**


α*, correlate; T, Tesla; ROI, Region of Interest; FA, fractional anisotropy; MD, Mean Diffusivity; AxD, Axial diffusivity; RD, Radial diffusivity. VBM, voxel based morphometry; TBSS, Tract based spatial statistics.*

At birth, the fornix is more prominent compared to the other brain fiber tracts and this phenotype is retained during infancy (Hermoye et al., 2006; Dubois et al., 2008). The development of fornix is thought to be completed by age 5 years (Hermoye et al., 2006; Dubois et al., 2008; Lebel et al., 2012) (**Figure 1B**). However, three cross-sectional (Lebel et al., 2012; Sala et al., 2012; Douet et al., 2014) and two longitudinal (Simmonds et al., 2014) DTI studies that investigated the volume and/or white matter integrity and density of the fornix showed its development through adolescence, and further age-related changes of the fornix throughout the lifespan (**Figure 1C**). Forniceal (body/column and crescent) FA exhibits an inverted U-shaped curve while the MD shows a U-shaped curve, and both peak at late adolescence (maximum at 19.5 years for FA and minimum at 17.8 years for MD) (Lebel et al., 2012; Sala et al., 2012). A recent longitudinal study reported no significant changes in the developmental trajectory of FA of the forniceal body/column, while FA in its crescent portion continues to increase during adolescence [13–16 years] (Simmonds et al., 2014). These findings suggest that the agerelated changes of FA observed in the cross-sectional studies were primarily due to changes in the crescent rather than in the body/column of the fornix. Before peaking at late adolescence, the fornix has the steepest age-dependent increase in MD amongst all major tracts (Lebel et al., 2012; Sala et al., 2012), with AxD and RD showing more than 2% change per year (Simmonds et al., 2014). Interestingly, age-related increase of AxD, i.e., accelerated prunning, was found in the left hemisphere but not in the right hemisphere during childhood and adolescence (Simmonds et al., 2014). Similarly, asymmetric atrophy of the hippocampus and fornix were reported in several neurological disorders such as schizophrenia (Crow et al., 1989; DeLisi et al., 1997; McDonald et al., 2000; Chance et al., 2005; Mitchell and Crow, 2005; Mitelman et al., 2005), bipolar disorders (Brisch et al., 2008), temporal lobe epilepsy (Baldwin et al., 1994; Hori, 1995; Kim et al., 1995; Kuzniecky et al., 1999), and in some patients with traumatic brain injury (Tate and Bigler, 2000; Tomaiuolo et al., 2004). Therefore, finding a more sensitive neuroimaging marker to assess the forniceal lateralization, such as AxD of the fornix, may be useful for early diagnosis of these disorders.

Prior to adulthood, the forniceal volume also exhibits an inverted U-shaped curve with age, and thereafter an agedependent decrease in the volume in both longitudinal and cross-sectional studies (Pagani et al., 2008; Giorgio et al., 2010; Michielse et al., 2010; Lebel et al., 2012; Sala et al., 2012; Fletcher et al., 2013).

During adulthood, the white matter integrity and density of the fornix typically decrease with age across DTI studies (Stadlbauer et al., 2008; Lee et al., 2009; Yasmin et al., 2009; Michielse et al., 2010; Sullivan et al., 2010; Lebel et al., 2012; Sala et al., 2012; Fletcher et al., 2013; Sasson et al., 2013). The majority of the studies found age-related decreases of the forniceal FA (Stadlbauer et al., 2008; Yasmin et al., 2009; Zahr et al., 2009; Burzynska et al., 2010; Michielse et al., 2010; Jang et al., 2011; Metzler-Baddeley et al., 2011; Lebel et al., 2012; Sala et al., 2012; Fletcher et al., 2013; Pelletier et al., 2013; Sasson et al., 2013), and only two studies showed no changes with age (Lee et al., 2009; Sullivan et al., 2010). Findings on the diffusivities (MD, AxD and RD) are less consistent and varied depending on the region of interest (crus, body/column or the entire fornix). While the majority of the studies found diffusivities (MD, AxD and RD) of the fornix increase with age (Stadlbauer et al., 2008; Yasmin et al., 2009; Zahr et al., 2009; Burzynska et al., 2010; Sullivan et al., 2010; Jang et al., 2011; Lebel et al., 2012; Sala et al., 2012; Sasson et al., 2013; Simmonds et al., 2014), several studies reported either age-related decrease of MD and AxD (Michielse et al., 2010) or no changes in the fornix with age for MD (Lee et al., 2009) and RD (Michielse et al., 2010) across the age span of 10 to 80 years. White matter maturation follows sex-specific differential trajectories (Westerhausen et al., 2004; Schmithorst et al., 2008; Asato et al., 2010). Girls showed maturation of white matter integrity earlier than boys (Asato et al., 2010). In particular, girls showed greater age-dependent increase of MD in associative regions compared to boys. Furthermore, tendencies for age-related increase of FA were found in the right hemisphere for girls but in the left hemisphere for boys. These sex-specific brain differences parallel the pubertal changes that occur during adolescence, suggesting that hormonal changes might influence white matter maturity. However, the few studies that examined the relationships between physical pubertal maturity and circulating hormones on white matter maturation were underpowered by sample size(Peper et al., 2008, 2009). Nevertheless, discrepancies on diffusivities between DTI studies are not due to differences in age range and sex distribution, since they are similar across all of these studies. The image processing methods for these studies are also similar between those that showed conflicting results. Therefore, sample size or inter-subject variations might have contributed to the different findings regarding the age-related changes in FA and MD.

Overall, the fornix is one of the earliest white matter tracts to mature. After its maturation peaks during late adolescence, the fornix begins to "atrophy" throughout the remainder of the lifespan. However, "pruning" rather than degenerative processes likely contribute to the early decreases in forniceal volume. More detail anatomical assessments of the fornix (column, body, and crus) and more systematic evaluations across a larger age range, followed longitudinally, are needed to better characterize the developmental trajectories of the fornix.

## *Relationship with cognition*

Fibers from the fornix comprise the main cholinergic input to the hippocampi and major efferent pathways from the hippocampi to the anterior thalamic nuclei, mammillary bodies, striata, and prefrontal cortices. These anatomical connections are involved in memory networks, which demonstrate that the fornix plays a critical and central role in memory tasks, particularly episodic memory. However, few studies investigated the relationships between forniceal metrics and memory tasks during typical development and aging. During childhood [3–20 years], larger forniceal volume was correlated with better episodic memory scores in healthy children. But this relationship was reversed in those carrying the *NRG1*-T-risk alleles for schizophrenia and psychosis (Douet et al., 2014). During young adulthood [22–31 years], greater FA in the fornix was associated with better episodic memory scores, especially with spatial recognition (Rudebeck et al., 2009). Similarly, across studies of young adults and older adults, forniceal FA correlated positively with working memory (Zahr et al., 2009), episodic memory (Rudebeck et al., 2009), and with both verbal and visual recall tasks (Rudebeck et al., 2009; Zahr et al., 2009; Metzler-Baddeley et al., 2011). A 4-year longitudinal follow-up study of healthy older adults found that lower forniceal volume and higher AxD at baseline predicted conversion to cognitive impairments (mild cognitive impairment or dementia) (Fletcher et al., 2013). Therefore, volumetric and white matter changes of the fornix appear to be effective biomarkers to validate or corroborate with memory performance across the lifespan, and to predict hippocampal function (Aggleton et al., 2000; Rudebeck et al., 2009; Fletcher et al., 2013; Pelletier et al., 2013).

#### **FORNIX AS A PREDICTOR OF MEMORY DEFICITS**

Early studies in humans did not report associated memory deficits after lesion of the fornix (Garcia-Bengochea and Friedman, 1987). More recent studies, however, consistently reported deficits in several cognitive abilities, especially in episodic memory, in patients with injuries to the fornix (Gaffan et al., 1991; Squire and Zola-Morgan, 1991; Aggleton et al., 2000). Moreover, as part of the limbic system, fornix degeneration may precede hippocampal dysfunction, and may predict conversion to cognitive impairment better than hippocampal atrophy (Fletcher et al., 2013). Hence, assessments of the fornix have recently become a major research focus in determining its role in neurological disorders that are associated with memory impairments.

## *Alzheimer disease and dementia syndromes (Table 2)*

Alzheimer disease (AD) and mild cognitive impairment (MCI) can be distinguished from normal aging by the different clinical syndromes (Petersen et al., 2001). MCI includes amnestic MCI (aMCI) and non-amnestic MCI (naMCI) (Petersen, 2004), depending on the memory impairment features. While naMCI patients tend to develop frontotemporal dementia or other types of dementias with broader cognitive deficits, aMCI patients are at risk for Alzheimer's disease (Mielke et al., 2014). In the US population, the prevalence of MCI ranges from 3 to 19% depending on the studies. About 40% of MCI patients will develop AD or other dementias, while most of MCI patients stay stable, and some even revert to a healthy control diagnosis (Mielke et al., 2014). Therefore, understanding prodromal AD and predicting accurately when MCI will convert to dementia can lead to early diagnosis and prevention of dementia when effective preventive strategies become available.

The neuropathology of AD is characterized by the presence of extracellular beta-amyloid plaques and intracellular neurofibrillary tangles that both lead to neuronal dysfunction and apoptosis (Bossy-Wetzel et al., 2004). Neurofibrillary tangles result from the intracellular oligomerization of the microtubule-associated protein Tau. The deposition of neurofibrillary tangles begins primarily in the limbic system structures, initially in the entorhinal cortex and the medial temporal regions, then progressively spread across the cerebral cortex. Hippocampal and entorhinal cortical atrophy assessed with MRI is well documented in patients with AD (Teipel et al., 2013), and in many with MCI (Pihlajamaki et al., 2009). Furthermore, this observation has extended the investigation of all limbic structures in relation to disease progression and cognitive performance.

The fornix is atrophied in MCI and AD patients compared to healthy controls, (Callen et al., 2001; Copenhaver et al., 2006; Ringman et al., 2007; Hattori et al., 2012) as confirmed by a longitudinal follow-up study (Douaud et al., 2013). Furthermore, in a large cohort of 79 aMCI and 204 healthy controls (HC), the volume of the crus of the fornix more specifically discriminated between MCI and HC (Cui et al., 2012).

Decreased FA of the fornix, on DTI, was found to be more sensitive than decreases in volume and/or area, on structural MRI, for predicting AD progression, since decreased FA preceded the atrophy more than two years prior to conversion from MCI to AD (Douaud et al., 2013). AD patients had lower FA (Liu et al., 2011b; Metzler-Baddeley et al., 2012) and higher MD and RD in the fornix compared to healthy controls (Mielke et al., 2009; Stricker et al., 2009; Liu et al., 2011b; Hattori et al., 2012; Huang et al., 2012; Oishi et al., 2012; Nowrangi et al., 2013; Zhuang et al., 2013), and at disease onset as defined by comparison between MCI and/or early onset AD patients (Mielke et al., 2009; Zhuang et al., 2010; Liu et al., 2011b; Oishi et al., 2012; Canu et al., 2013; Douaud et al., 2013; Nowrangi et al., 2013). A similar phenotype of lower FA in the fornix was found also in patients with genetically inherited dementias in comparison to controls (Ringman et al., 2007). Longitudinal studies showed that the magnitude of age-related changes of DTI metrics is similar between AD, MCI and healthy controls (Mielke et al., 2009, 2012; Oishi et al., 2012), suggesting that abnormal forniceal FA and MD are likely to predict convertion from MCI to AD.

Lower FA and higher diffusivity metrics in the fornix were associated also with worse performance on short- and longterm memory tasks and with clinical dementia evaluations in AD and MCI patients (Ringman et al., 2007; Mielke et al., 2009, 2012; Kantarci et al., 2011; Zhuang et al., 2013), as well as in healthy controls (Sexton et al., 2011; Oishi et al., 2012; Nowrangi et al., 2013). These cognitive measures showed deficits in verbal memory (i.e., California Verbal Learning Test, Hopkins Verbal Learning Test) and visual memory (Rey-Osterrieth Complex Figure Test), as well as in more global measures (MMSE and Clinical Dementia Rating).

Therefore, measurements of macro- and micro-structural changes in the fornix may provide preclinical surrogate markers to predict the development of Alzheimer disease and allow early treatment in these patients.

#### *Schizophrenia (SCZ) (Table 3)*

Clinical signs, brain imaging and genetic studies all contributed to the hypothesis that schizophrenia and psychiatric diseases are neurodevelopmental disorders (Rapoport et al., 2012) with neurodegenerative components (Vita et al., 2012). In addition, findings from postmortem and neuroimaging studies suggest that white matter maturation and myelination processes are disrupted in schizophrenia, which might trigger its symptoms (Heckers et al., 1991; Arnold et al., 1995) or lead to age-related white matter loss and cognitive decline (Chang et al., 2007; Kochunov and Hong, 2014). Brain abnormalities in SCZ patients occur in the paralimbic and temporolimbic regions (Kasai et al., 2003), which are involved in episodic memory. Incidentally, episodic memory impairment is one of the most consistent phenotype for schizophrenia (Schaefer et al., 2013). Since the fornix is part of the limbic system, and is involved in episodic memory, it has been evaluated with histopathology and brain imaging in SCZ patients.

Histopathologic studies showed that SCZ men, but not women, had greater than normal fiber density in the left fornix, suggesting sex and hemisphere specific alterations in the myelination of the fornix in schizophrenia (Chance et al., 1999). However, the fornix volume and cross-sectional area were found to be similar between SCZ patients and healthy adult controls in


#### **Table 2 | Forniceal macro- and micro-structure alterations in patients with Alzheimer's disease and mild cognitive impairments.**


#### **Table 2 | Continued**

*MRI, Magnetic Resonance Imaging; T, Tesla; TBSS, Tract-based spatial statistic; VBA, Voxel-based analysis; ROI, Region of Interest; FA, Fractional Anisotropy; MD, Mean Diffusivity, AxD, Axial Diffusivity; RD, Radial Diffusivity.*

*EOAD, early-onset Alzheimer's disease; LOAD, late-onset Alzheimer's disease; HC, Healthy controls; naMCI, non-amnesic; MCI;AD, Alzheimer disease; iNPH, idiopathic normal pressure hydrocephalus; FAD, familial Alzheimer's disease; FADmc, familial Alzheimer's disease mutation carriers, FADnc, familial Alzheimer's disease non-carriers; pMCI, amnestic MCI patients who progressed to probable AD no earlier than 2 years after their baseline scan; sMCI, amnestic MCI patients who were clinically stable i.e., did not develop AD for at least 3 years following their first evaluation.*

*WMS–R, Wechsler Memory Scale–Revised; CDR, Clinical Dementia Rating; RCFT, Rey Complex Fig Test; HVLT-R, Hopkins Verbal Learning Test-Revised.*


#### **Table 3 | Forniceal macro- and micro-structure alterations in schizophrenia and psychiatric disorders.**

#### **Table 3 | Continued**


α*: correlate.*

*MRI, Magnetic Resonance Imaging; T, Tesla; TBSS, Tract-based spatial statistic; VBA, Voxel-based analysis; ROI, Region of Interest; FA, Fractional Anisotropy; MD, Mean Diffusivity, AxD, Axial Diffusivity; RD, Radial Diffusivity.*

*HC, Healthy controls; SCZ, schizophrenic patients; PsyC, psychiatric controls non–schizophrenics: ADHD, Attention deficit hyperactivity disorder; uDep, unipolar Depression, bDep, bipolar depression; BPD, bipolar disorder; FES, first episode schizophrenia.*

*DPT, Doors and People Test.*

postmortem brain tissues (Brisch et al., 2008), and in an *in vivo* MRI study (Zahajszky et al., 2001). In contrast, larger fornices were found on MRI of adolescent SCZ (ages 16–17 years, both males and females) compared to healthy controls and to patients with other serious psychiatric disorders (Davies et al., 2001). These variable findings regarding the forniceal volume might have resulted from the different subject populations and the less well defined fornix structures on these earlier structural MRI studies.

Findings on the fornix measurement have been more consistent across DTI studies. Using tractography, forniceal bundle volume in SCZ adolescents and adults were smaller [−11–16%] than in healthy controls (Kuroki et al., 2006; Kendi et al., 2008). The various DTI studies and approaches, using tractography, regions of interest (ROI) and tract-based spatial statistics (TBSS), consistently showed that FA of the fornix is lower in SCZ patient compared to healthy control. The lower than normal FA appears early at the onset of SCZ, which typically occurs just before adolescence (Davenport et al., 2010). This phenotype was reported in adolescent patients with their first episode of SCZ (Lee et al., 2009), in SCZ young adults (Luck et al., 2010; Fitzsimmons et al., 2014) and in mid-life SCZadults (Kuroki et al., 2006; Takei et al., 2008; Fitzsimmons et al., 2009; Abdul-Rahman et al., 2011), suggesting that lower than "normal" FA is a stable marker for SCZ that is retained throughout the lifespan. Lower FA in SCZ patients is frequently accompanied by either higher MD (Kuroki et al., 2006; Takei et al., 2008), RD (Abdul-Rahman et al., 2011) or both (Fitzsimmons et al., 2014). Findings on AxD are less consistent. Two studies showed no changes in AxD between SCZ patients and healthy controls (Kendi et al., 2008; Abdul-Rahman et al., 2011), whereas another study found higher AxD along with higher MD and RD in young adults with first episode schizophrenia (Fitzsimmons et al., 2014). The higher RD was suggested to be a marker of myelin disruption, higher MD a marker of atrophy, while AxD may reflect axonal disruption (Song et al., 2005). Therefore, the lower FA and higher diffusivities involving all three measures (MD, RD and AxD) possibly reflect alterations in both myelin and axons. These alterations are notable in the fornix already at illness onset, but the causative mechanism is not yet defined. Moreover, the functionality of the forniceal changes and their impacts on the limbic network is still unclear. Some studies found no association between forniceal metrics and either cognitive or clinical measures (Zahajszky et al., 2001; Fitzsimmons et al., 2009, 2014; Lee et al., 2013), while others reported that lower FA correlated with greater psychopathology (Abdul-Rahman et al., 2011) and higher medication dosage in SCZ patients (Kuroki et al., 2006). Lower FA and/or higher MD was further associated with greater episodic memory impairments (verbal and visual memory tests) in SCZ patients (Nestor et al., 2007; Takei et al., 2008; Lee et al., 2013). In healthy controls, these correlations between FA and visual and verbal memory tasks were also observed (Nestor et al., 2007; Fitzsimmons et al., 2009).

In conclusion, abnormalities in the fornix are found in SCZ patients and are most likely due to degeneration, involving both axonal injury and demyelination, of the fornix. To some extent, these microstructural abnormalities in the fornix may serve as an imaging marker for disease severity in schizophrenia, although it remains unclear whether these changes in the fornix contribute to disruption of the limbic networks and to hippocampal atrophy. In addition, DTI metrics (FA and MD) appear to be sensitive indicators of injury to the fornix and subsequent memory deficits in SCZ patients. Further investigations using these metrics, in addition to other imaging modalities (e.g., evaluating brain network connectivities), are needed to follow patients longitudinally from the prodromal period to the first episodes to understand further the evolution of the neuropathology of schizophrenia.

## *Multiple Sclerosis and other neurodegenerative diseases (Table 4)*

Multiple Sclerosis (MS) is an autoimmune demyelinating disease that is characterized by the infiltration of macrophages and T-cells that activate glia and microglia, which lead to fulminant neuroinflammation and intense demyelination of nerve fibers (Pivneva, 2008). About half of the MS patients develop cognitive deficits and most frequently, episodic memory deficits (Brissart et al., 2011). As parts of the limbic system, both the hippocampus and the fornix were often found affected in MS patients. Compared to healthy controls, MS patients had lower magnetization transfer ratio (MTR) in the right fornix, but this abnormality in the fornix did not correlate with cognitive performance (Ranjeva et al., 2005). Using TBSS, tractography or ROI, MS patients consistently showed lower FA with higher MD and RD in the fornix than healthy controls across studies and during adulthood (Smith et al., 2006; Dineen et al., 2009, 2012; Roosendaal et al., 2009; Fink et al., 2010; Kern et al., 2012; Koenig et al., 2013; Syc et al., 2013). Findings on AxD in the fornix were less consistent and less systematically investigated. Forniceal AxD showed either no group differences (Dineen et al., 2012) or higher values in MS compared to healthy controls (Roosendaal et al., 2009; Syc et al., 2013). In most of these studies, MS patients with lower FA and higher diffusivity metrics in the fornix had poorer performance in verbal and visual memory or recall and greater episodic memory impairments (Brief Visual Memory Test-Revised) (Dineen et al., 2009, 2012; Koenig et al., 2013, 2014; Syc et al., 2013). Moreover, these forniceal DTI metrics correlated with Expanded Disability Status Scale (EDSS) and disease duration in these MS patients (Syc et al., 2013; Koenig et al., 2014).

Altogether, these findings showed that DTI metrics in the fornix are consistently abnormal in MS patients. Since DTI measures in the fornix can assess disease severity, they may be useful for monitoring MS disease progression. Furthermore, forniceal DTI metrics correlated with hippocampal volume in patients with MS, and DTI measures in the fornix had an even stronger association with visual and episodic memory than the hippocampal volume (Koenig et al., 2014). Therefore, similar to patients who convert from MCI to AD, longitudinal forniceal DTI measures may be useful in predicting hippocampal abnormalities and memory deficits in MS patients.

*Parkinson's disease.* Parkinson's disease (PD) is most commonly regarded as a movement disorder (Gelb et al., 1999), since degeneration of the nigrostriatal dopaminergic system leads to dysfunction of the motor system with the four cardinal signs of tremors, bradykinesia, rigidity and postural instability. However, dopamine also mediates attention and working memory, which are requried for most higher level cognitive function, Therefore, PD patients commonly develop dementia and cognitive deficits including deficits in executive function, attention, language and memory (Zgaljardic et al., 2003; McKinlay et al., 2010). Few studies investigated the fornix in PD patients using DTI. Similar to AD patients, MD of the fornix was higher in PD patients than in healthy adults (Kim et al., 2013), and higher MD correlated with worse short-term non-verbal memory (Zheng et al., 2014). However, since many dementia patients have co-occurrence of AD and PD, it remains unclear whether the fornix abnormalities are related specifically to PD. Depression is also common amongst PD patients, and those with depression showed lower FA in the frontal white matter than PD patients without depression; although the fornix was not specifically evaluated in this study, and the temporal white matter showed no group difference (Matsui et al., 2007). Another DTI study found that PD patients with excessive daytime sleepiness (Epsworth Sleepiness Scale ≥ 10) had significantly lower FA in their fornix compared to controls (Matsui et al., 2006). Therefore, abnormalities in the fornix appear to contribute to the co-morbid symptoms beyond the extrapyramidal system, such as memory deficits and excessive daytime sleepiness in patients with PD.

*Epilepsy.* Since a large number of etiologies exist for epilepsy, the fornix may or may not be affected depending on whether this major efferent white matter tract from the hippocampi is affected by the lesion or condition that caused the epilepsy. For instance, mesial temporal sclerosis (MTS) causes temporal lobe epilepsy (TLE), and is frequently accompanied by forniceal atrophy and lower FA when the fornix of these patients are compared to non-epileptic controls (Baldwin et al., 1994; Kim et al., 1995). Decreased fornix volumes and lower FA were often associated with ipsilateral hippocampal sclerosis, both quantitatively and qualitatively (Baldwin et al., 1994; Kuzniecky et al.,


#### **Table 4 | Forniceal macro- and micro-structure alterations in multiple sclerosis and other neurodegenerative diseases.**

*(Continued)*

#### **Table 4 | Continued**


*(Continued)*

#### **Table 4 | Continued**


#### α*: correlate.*

*MRI, Magnetic Resonance Imaging; T, Tesla; MTR, magnetization transfer ratio; TBSS, Tract-based spatial statistic; VBA, Voxel-based analysis; ROI, Region of Interest; MO, mode of anisotropy; FA, fractional anisotropy; MD, Mean Diffusivity, AxD, Axial diffusivity; RD, Radial diffusivity.*

*MS, Multiple sclerosis; PD, Parkinson's disease; RRMS, relapsing-remitting multiple sclerosis; SPMS, secondary progressive multiple sclerosis; PPMS, primary progressive multiple sclerosis; CISSMS, clinically isolated syndrome suggestive of multiple sclerosis; JME, juvenile myoclonic epilepsy; IGE-GTC, generalized tonic– clonic seizures; MTLE, Mesial temporal lobe epilepsy; MS, Mesial Temporal Sclerosis; preHS, pre-surgical hippocampal sclerosis; postHS, post-surgical hippocampal sclerosis; TBI,Traumatic brain injury.*

*EDSS, Expanded Disability Status Scale; 9-HPT, 9-Hole Peg Test; PASAT-3, Paced Auditory serial Addition Task-3, second version; GMI, General Memory Index score; WMS–R, Wechsler Memory Scale–Revised; RCFT, Rey Complex Fig Test; BVMT-R, Brief Visuospatial Memory Test-Revised; SDMT, Symbol Digit Modalities Test.*

1999), and appear to be a good predictor of TLE with accurate lateralization. Therefore, evaluating the fornix and its asymmetry, even with visual interpretations, may be useful in support of presurgical planning (i.e., for surgical resection) for patients with medically intractable TLE. In addition, low frequency depth electrode stimulation of the fornix led to hippocampal and posterior cingulum responses, demonstrating these functional connections, and reduced interictal epileptiform discharges and seizures in patients with intractable mesial temporal lobe epilepsy, without affecting their memory (Koubeissi et al., 2013). Depth electrode stimulation of the fornix also led to either ipsilateral or contralateral hippocampal responses, which again confirmed these neural pathway connections and explained how seizure discharge might spread between homotopic mesial temporal structures without neocortical involvement (Lacuey et al., 2014).

Lastly, a recent DTI study, using tractography, found that patients with juvenile myoclonic epilepsy (JME) had lower FA in the crus of the fornix, body of the corpus callosum and many other major white matter tracts, but not in those with only generalized tonic-clonic seizures, suggesting different neuroanatomical substrates in these two different types of idiopathic generalized epilepsies (Liu et al., 2011a). Taken together, these studies demonstrate that the fornix may play a role in mediating seizure spreads across the cerebral hemispheres both in patients with temporal lobe epilepsy as well as in generalized epilepsies. However, it may also be a treatment target for deep brain stimulation or surgical approaches in these patients.

*Traumatic Brain Injury (TBI)* results from physical forces that damage the brain, which may cause cognitive impairments such as memory and attention deficits. TBI is also associated with atrophy of the fornix (Gale et al., 1993; Tate and Bigler, 2000; Tomaiuolo et al., 2004). The fornix is particularly susceptible to physical shearing forces (Tate and Bigler, 2000), probably due to its delicate fiber tracts that straddle both cerebral hemispheres. The effects of TBI on the fornix volume have been examined in relation to memory. However, only one of the three studies (Tate and Bigler, 2000; Gale et al., 1993; Tomaiuolo et al., 2004) found a correlation between forniceal atrophy and memory performance (Tomaiuolo et al., 2004). However, using DTI with TBSS, white matter abnormalities were observed in several regions of the brain in TBI patients, but only lower FA in the fornix correlated with worse performance in associative memory and learning in both the TBI and healthy control groups (Kinnunen et al., 2011). In another study, patients with diffuse TBI, which leads to diffuse axonal injury, had globally decreased FA in the brain. However, regional analyses showed that lower FA in the superior longitudinal fasciculus was associated with working memory deficits, while lower forniceal FA was associated with poorer declarative memory in these TBI patients with diffuse injuries (Palacios et al., 2011). Lower forniceal FA and memory deficits were consistently found in TBI patients compared to healthy controls (Palacios et al., 2011; Adnan et al., 2013), suggesting a critical role of the fornix integrity in the development of memory impairments after TBI.

# **DISCUSSION**

#### **LIMITATIONS AND FUTURE STUDIES**

Several factors have limited the study of the fornix. First, the anatomy of the fornix makes it difficult to evaluate the abnormalities in this brain structure. Specific regions of the fornix (i.e., column, crus or pre-commissural fornix) are even more difficult to visualize or quantify. DTI has improved the visualization of the fornix, which has led to many more studies of this structure in several neurological disorders. However, the forniceal DTI measures in the published studies are often affected by the fornix's close proximity to the ventricles, which can lead to partial volume effects from the CSF in the ventricles. While CSF suppression (using inversion recovery pulses on MRI) would suppress the signals from CSF, most of the DTI studies did not apply such CSF signal suppression during the image acquisition. Partial volume effect from adjacent CSF signal may generate biased (higher) diffusivity and (lower) FA values of the fornix although this structure is generally well delineated on DTI. Nevertheless, refined methods have been developed to minimize the CSF partial volume effect on DTI and obtain higher resolution images. These improved DTI acquisition methods include reducing the repetition time and using non-zero minimum diffusion weighting (Baron and Beaulieu, 2014), or increased the image resolution of DTI (Herbst et al., 2014) by combining multiplexed sensitivity encoding (Chen et al., 2013) and prospective motion correction (Zaitsev et al., 2006; Herbst et al., 2012; Gumus et al., 2014). Others have developed novel criteria for DTI metric selection (Pasternak et al., 2010) using relationships between distribution and distance of the measured diffusion quantities or the use of multi-contrast MRI (Tang et al., 2014) with an automated parcellation atlas, which may further delineate and accurately assess how the fornix might be affected in various brain disorders. Second, the various methods (e.g., manual or automated ROIs, TBSS, tractography) used to measure the diffusivities and FA in the fornix do not always yield the same results. Systematic comparisons or the use of more than one approach to measure the fornix in the same datasets could provide validation to the abnormal findings in the various brain disorders. Third, the majority of the studies reviewed utilized a cross-sectional design, longitudinal follow-up studies would minimize the potential confounding effects of inter-subject variability (e.g., due to differences in disease severity or illness duration) or premorbid group differences. Intra-subject measurements in longitudinal evaluations are more sensitive in detecting, predicting and monitoring neurodegeneration compared to cross-sectional measures. Despite these limitations, the studies reviewed consistently showed correlations between DTI metrics in the fornix and memory performance of typically aging individuals and in patients with various neurodegenerative and neuropsychiatric disorders. These findings strengthen the role of the fornix as a useful imaging marker to predict memory deficits or impairments.

#### **CONCLUSION**

The fornix is clearly a critical component of the limbic system and is closely linked to memory performance. Alterations of the fornix are related to cognitive functions in childhood and in later life. In addition, forniceal changes were found in schizophrenia and other psychiatric disorders. Therefore, the fornix appears to be more than a clinical surrogate marker of memory impairments for neurodegenerative and neuroinflammatory diseases, such as Alzheimer's disease and multiple sclerosis. Although the fornix is one of the less heritable brain structures (Jahanshad et al., 2013), few studies reported how genes might influence the typical development or aging of the fornix. Imaging genetics might be useful to further elucidate the role of the fornix in various brain disorders as well as during healthy neurodevelopment and brain aging.

#### **ACKNOWLEDGMENTS**

We thank Eric Cunningham and Kristin Lee for carefully editing this manuscript. NIH grant supports: K24 DA016170 & G12-MD007601.

#### **REFERENCES**


a diffusion MRI tractography study. *J. Neurosci*. 31, 13236–13245. doi: 10.1523/JNEUROSCI.2317-11.2011


correction using an external optical motion tracking system. *Neuroimage* 31, 1038–1050. doi: 10.1016/j.neuroimage.2006.01.039


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 18 July 2014; accepted: 14 December 2014; published online: 14 January 2015.*

*Citation: Douet V and Chang L (2015) Fornix as an imaging marker for episodic memory deficits in healthy aging and in various neurological disorders. Front. Aging Neurosci. 6:343. doi: 10.3389/fnagi.2014.00343*

*This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2015 Douet and Chang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Fornix white matter is correlated with resting-state functional connectivity of the thalamus and hippocampus in healthy aging but not in mild cognitive impairment – a preliminary study

#### **Elizabeth G. Kehoe<sup>1</sup>\*, Dervla Farrell <sup>1</sup> , Claudia Metzler-Baddeley <sup>2</sup> , Brian A. Lawlor <sup>3</sup> , Rose Anne Kenny <sup>4</sup> , Declan Lyons <sup>5</sup> , Jonathan P. McNulty <sup>6</sup> , Paul G. Mullins <sup>7</sup> , Damien Coyle<sup>8</sup> and Arun L. Bokde<sup>1</sup>**

<sup>1</sup> Trinity College Institute of Neuroscience and Cognitive Systems Group, Discipline of Psychiatry, School of Medicine, Trinity College Dublin, Dublin, Ireland <sup>2</sup> Cardiff University Brain Research Imaging Centre (CUBRIC), Neuroscience and Mental Health Research Institute (NMHRI), School of Psychology, Cardiff University,

Cardiff, UK <sup>3</sup> Department of Psychiatry, Jonathan Swift Clinic, St. James Hospital, Trinity College Institute of Neuroscience, Trinity College Dublin, Dublin, Ireland


8 Intelligent Systems Research Centre, University of Ulster, Derry, UK

#### **Edited by:**

Kenichi Oishi, Johns Hopkins University, USA

#### **Reviewed by:**

Kenichi Oishi, Johns Hopkins University, USA Andreia Vadconcellos Faria, Johns Hopkins University, USA Kentaro Akazawa, Kyoto Prefectural University of Medicine, Japan

#### **\*Correspondence:**

Elizabeth G. Kehoe, Trinity College Institute of Neuroscience, Lloyd Institute, Room 3.36b, Trinity College Dublin, Dublin 2, Ireland e-mail: elkehoe@tcd.ie

In this study, we wished to examine the relationship between the structural connectivity of the fornix, a white matter (WM) tract in the limbic system, which is affected in amnestic mild cognitive impairment (aMCI) and Alzheimer's disease, and the resting-state functional connectivity (FC) of two key related subcortical structures, the thalamus, and hippocampus.Twenty-two older healthy controls (HC) and 18 older adults with aMCI underwent multi-modal MRI scanning.The fornix was reconstructed using constrained-spherical deconvolution-based tractography. The FC between the thalamus and hippocampus was calculated using a region-of-interest approach from which the mean time series were exacted and correlated. Diffusion tensor imaging measures of the WM microstructure of the fornix were correlated against the Fisher Z correlation values from the FC analysis.There was no difference between the groups in the fornix WM measures, nor in the resting-state FC of the thalamus and hippocampus. We did however find that the relationship between functional and structural connectivity differed significantly between the groups. In the HCs, there was a significant positive association between linear diffusion (CL) in the fornix and the FC of the thalamus and hippocampus, however, there was no relationship between these measures in the aMCI group. These preliminary findings suggest that in aMCI, the relationship between the functional and structural connectivity of regions of the limbic system may be significantly altered compared to healthy ageing.The combined use of diffusion weighted imaging and functional MRI may advance our understanding of neural network changes in aMCI, and elucidate subtle changes in the relationship between structural and functional brain networks.

**Keywords: diffusion MRI, tractography, functional connectivity, fornix, mild cognitive impairment (MCI), hippocampus, thalamus**

#### **INTRODUCTION**

Alzheimer's disease (AD) is the most common cause of neurodegenerative dementia and is often preceded by a stage known as amnestic mild cognitive impairment (aMCI) (Petersen et al., 1999; Petersen, 2004), which confers an increased risk of developing AD (Stephan et al., 2012). With the prevalence of AD predicted to rise substantially over the coming years (Barnes and Yaffe, 2011), there has been increasing interest in using neuroimaging to understand early brain changes associated with preclinical groups such as aMCI (Jack et al., 2010; Sperling et al., 2011), which may represent the earliest stages of the disease. Early neuroimaging studies of

aMCI and AD predominantly focused on localized brain changes such as hippocampal atrophy (Hua et al., 2010; Jack et al., 2010), however there is now a large body of evidence suggesting that the brain's inherent structural and functional connectivity (FC) is disrupted in aMCI and AD (Seeley et al., 2009; Teipel et al., 2014; Vidal-Pineiro et al., 2014).

Functional connectivity studies have found reduced connectivity in resting-state networks in aMCI and AD, including the default mode network (DMN) (Greicius et al., 2004; Beason-Held, 2011; Brier et al., 2012; Jacobs et al., 2013; Sheline and Raichle, 2013), and have found that altered FC is also linked to memory

<sup>4</sup> Mercer's Institute for Successful Ageing, St. James Hospital, Trinity College Institute of Neuroscience, Trinity College Dublin, Dublin, Ireland

dysfunction in these populations (Wu et al., 2013; Dunn et al., 2014). Studies using diffusion weighted imaging (DWI) and tractography to measure white matter (WM) connectivity in aMCI and AD have found significant alterations in tracts essential for memory function such as the fornix, uncinate fasciculus, and cingulum (Naggara et al., 2006; Muller et al., 2007; Stahl et al., 2007; Catheline et al., 2010; Fellgiebel and Yakushev, 2011; Metzler-Baddeley et al., 2012a,b).

The fornix is the main efferent pathway of the hippocampus, connecting it with regions including the mammillary bodies of the hypothalamus and the thalamus (Schmahmann and Pandya, 2006; Aggleton et al., 2010). A number of studies have found changes in the fornix in aMCI and AD (Mielke et al., 2009; Zhuang et al., 2010; Bozoki et al., 2012; Lee et al., 2012), such as lower fractional anisotropy (FA – thought to reflect fiber tract density and myelination). Furthermore, in a prospective study in healthy older adults, lower FA in the fornix at baseline predicted conversion to aMCI after 2 years (Zhuang et al., 2012a).

Although memory dysfunction in aMCI has been linked with both hippocampal atrophy (Convit et al., 1997; Grundman et al., 2003; Stoub et al., 2006) and WM degeneration in the fornix (Zhuang et al., 2012b), however, the role of the thalamus in aMCI is less well-defined, even though this structure is closely associated with the fornix and hippocampus. The thalamus plays a major role in generating the many rhythms in electroencephalography (EEG), which change substantively during neurodegeneration (Cantero et al., 2009); yet little is known about the role of the thalamus in neurodegeneration, and whether or not thalamus atrophy is a primary or secondary phenomenon to hippocampal or cortical atrophy in AD.

Several volumetric studies have found that the thalamus does undergo neurodegeneration in aMCI and AD (Chetelat et al., 2005; Shiino et al., 2006; de Jong et al., 2008; Cherubini et al., 2010; Roh et al., 2011; Zhang et al., 2013), and in a study that combined shape analysis of the thalamus and diffusion tensor imaging (DTI) (Zarei et al., 2010), thalamic regions most highly connected to the hippocampus showed the most severe atrophy in aMCI. Changes in the FC of the thalamus have been reported in aMCI (Zhou et al., 2013) albeit less frequently than changes in hippocampal FC, and in one study of healthy elders (Ystad et al., 2010), a negative correlation was found between thalamic FC and verbal free recall. This suggested that higher performers displayed more de-synchronization of thalamic signals, and that the FC of the thalamus may be linked to memory function. Yoon et al. (2012) found a positive correlation between perfusion in the left thalamus and performance on the Rey complex figure test in MCI, suggesting a role for the thalamus in cognitive decline.

In the current study, we used DWI and resting-state functional MRI (fMRI) to examine the connectivity of the fornix, thalamus, and hippocampus in aMCI, putatively viewed as a set of connected brain structures, which together form part of a limbic episodic memory network that malfunctions if one or more components are impaired. In particular, we wished to examine whether there was a correspondence between the structural and FC measures, and whether this was altered in aMCI. This is the first study to our knowledge to examine the link between structural and FC of the thalamus and hippocampus in aMCI. Previous studies have suggested that functional and structural connectivity are closely related, particularly in healthy participants (Damoiseaux and Greicius, 2009; Greicius et al., 2009), however the link between the two is not well-defined (Sporns, 2014), especially in the context of ageing and aMCI. Disruption of anatomical connections may influence the organization of FC, and combining MRI modalities should provide greater insight into the neural network connectivity changes in aMCI.

We performed constrained-spherical deconvolution (CSD) based fiber tractography of the fornix in a cohort of older adults with aMCI and healthy age-matched controls, and examined the resting-state FC of the thalamus and hippocampus. Taken together, these MRI measures should provide a fuller picture of the changes that occur in the fornix in aMCI, as well as about the relationships between microstructural WM changes and alterations in FC. We predicted that the aMCIs would show decreased FA (thought to reflect fiber tract density and myelination; Mori and Zhang, 2006), and increased mean diffusivity values (MD – which tends to be low in highly intact,organized tracts and which usually increases in disease states and neurodegeneration; Mori and Zhang, 2006) in the fornix relative to controls, in line with previous studies (Mielke et al., 2009; Zhuang et al., 2010; Bozoki et al., 2012; Lee et al., 2012). We also predicted that there would be reduced restingstate FC of the thalamus and hippocampus, similar to previous findings (Zhou et al., 2008, 2013), which may be related to WM microstructural alterations in the fornix.

#### **MATERIALS AND METHODS**

#### **PARTICIPANTS**

Twenty-two older healthy control (HC) participants and 19 older participants with aMCI took part in the study. The tractography analysis was unsuccessful for one aMCI participant, in whom the tracts were almost completely absent. An examination of their T1-weighted anatomical scan revealed quite advanced levels of atrophy and gross enlargement of the lateral ventricles. Therefore, 18 aMCI were included in the final sample.

The HCs were community-dwelling older adults recruited from the greater Dublin area *via* newspaper advertisements. They underwent a health screening questionnaire and a neuropsychological assessment, the Consortium to Establish a Registry for Alzheimer's Disease (CERAD , Morris et al., 1988), in order to rule out possible cognitive impairment before inclusion in the study. The CERAD battery has been shown to be sensitive to the presence of age-related cognitive decline (Welsh et al., 1991, 1992). All of the older participants included in the study scored no more than 1.5 SD below the standardized mean scores for subjects of a similar age and education level on any of the sub-tests.

The aMCI participants were recruited from memory clinics in St. James Hospital and St. Patrick's Hospital in Dublin, and were diagnosed by a clinician according to the Peterson criteria (Petersen et al., 1999) – i.e., abnormal memory scores for age and education level with no dementia. Four were single amnestic MCI (aMCI), and 14 were multi-domain aMCI (Petersen, 2004). Neuropsychological measures were administered or supervised by an experienced neuropsychologist and included the mini-mental state examination (MMSE; Folstein et al., 1975) and Cambridge cognitive examination (CAMCOG; Huppert et al., 1995).

All of the participants were right-handed with no history of head trauma, neurological disease, stroke, transient ischemic attack, heart attack, or psychiatric illness. They completed the Geriatric Depression Scale (GDS; Yesavage, 1988), the Eysenck Personality Questionnaire Revised Edition Short Scale (EPQ-R; Eysenck and Eysenck, 1991), and a Cognitive Reserve Questionnaire (Rami et al., 2011) before the MRI scan. The groups did not differ in terms of age, gender, education level, or levels of cognitive reserve as assessed by the self-report Cognitive Reserve Questionnaire. The aMCI group had lower MMSE scores, higher GDS scores, and scored lower on the EPQ measure of extraversion than the HC group. See **Table 1** for a summary of the participant demographics.

The study had full ethical approval from the St. James Hospital and the Adelaide and Meath Hospital, incorporating the National Children's Hospital Research Ethics Committee. All participants gave written informed consent before taking part in the study.

#### **MRI DATA ACQUISITION**

Whole-brain high angular resolution diffusion imaging (HARDI) data were acquired on a 3.0 Tesla Philips Intera MR system (Best, The Netherlands) equipped with an eight channel head coil. A parallel sensitivity encoding (SENSE) approach (Pruessmann et al., 1999) with a reduction factor of two was used during the DWI acquisition. Single-shot spin echo-planar imaging was used to acquire the DWI data with following parameters: Echo time (TE) 79 ms, repetition time (TR) 20,000 ms, field of view (FOV) 248 mm, matrix 112 × 112, isotropic voxel of 2.3 mm × 2.3 mm × 2.3 mm, and 65 slices with 2.3 mm thickness with no gap between the slices. Diffusion gradients were applied in 61 isotropically distributed orientations with *b* = 3000 s/mm<sup>2</sup> , and four images with *b* = 0 s/mm<sup>2</sup> were also acquired. The total scan time was 17 min.

A high-resolution 3D T1-weighted anatomical image was acquired for each participant with the following parameters: TE = 3.9 ms, TR = 8.5 ms, FOV = 230 mm, slice thickness = 0.9 mm,


MMSE, mini-mental state exam; GDS, geriatric depression scale; EPQ E, Eysenck personality questionnaire extraversion scale; EPQ N, Eysenck personality questionnaire neuroticism scale; CR, cognitive reserve scale. Standard deviations are indicated in parentheses.

\*Results of independent samples t-tests, except for gender which was compared with a Fischer's exact test. Statistically significant differences are indicated in bold font.

voxel size = 0.9 mm × 0.9 mm × 0.9 mm. These images were used for the correction of EPI-induced geometrical distortions in the DWI data.

Resting-state fMRI data were also acquired during the scanning session. The scan lasted for 7 min during which time the participants were asked to keep their eyes open and fixate on a cross hairs in the center of a screen behind the MR scanner, visible *via* a mirror. The blood oxygenation dependent (BOLD) signal changes were measured using a T2\*-weighted echo-planar imaging sequence with TE = 30 ms and TR = 2000 ms. Each volume of data covered the entire brain with 39 slices, and the slices were acquired in interleaved sequence from inferior to superior direction. Two hundred ten volumes of data were acquired, with voxel dimensions of 3.5 mm × 3.5 mm × 3.85 mm and a 0.35 mm gap between the slices.

#### **DWI ANALYSIS**

The DWI data were analyzed using ExploreDTI v4.8.3 (Leemans et al., 2009) 1 . The images were corrected for distortion due to head motion, eddy currents and for EPI-induced geometrical distortions by co-registration and resampling to the high-resolution T1-weighted anatomical images. This was implemented in ExploreDTI according to the method described by Irfanoglu et al. (2012) and the encoding vectors were reoriented appropriately (Leemans and Jones, 2009). Since ageing and neurodegeneration are related to brain atrophy and the fornix is extremely susceptible to contamination from cerebrospinal fluid (CSF) and atrophy-based partial volume artefacts, the free water elimination approach (Pasternak et al., 2009) was applied to correct for partial volume effects prior to fitting the tensor model to the data in each voxel. The free water elimination method has been used successfully in several previous tractography studies of ageing and aMCI, particularly in relation to the fornix (Metzler-Baddeley et al., 2012a,b; Fletcher et al., 2014).

#### **Tractography of the fornix**

For the analysis of the fornix a hybrid analysis approach was used, whereby reconstruction of the tracts was completed using CSDbased tractography (Jeurissen et al., 2011) and DTI-based WM indices were extracted for statistical analyses. CSD rather than DTI-based tractography was chosen as it can account for complex WM orientation such as crossing fibers (Tournier et al., 2008, 2011), and this approach has recently been successful at detecting changes in tracts with complex WM architecture in MCI and AD (Metzler-Baddeley et al., 2012b; Reijmer et al., 2012). In the case of the fornix for instance, DTI-based tractography cannot resolve complex fiber architecture in regions where the anterior columns of the fornix cross with fibers of the anterior commissure (Metzler-Baddeley et al., 2012a). Several previous studies have successfully used deterministic tractography based on the CSD method to segment the fornix (Metzler-Baddeley et al., 2011, 2013).

After the pre-processing steps whole-brain tractography was performed using every voxel as a seed point. The principle diffusion orientation at each point was estimated by the CSD tractography algorithm, which propagated in 0.5 mm steps along this

<sup>1</sup>http://www.exploredti.com/

direction. At each new location the fiber orientation(s) was estimated before the tracking moved a further 0.5 mm along the direction that subtended the smallest angle to the current trajectory. A trajectory was followed through the data until the scaled height of the fiber orientation density function peak dropped below 0.1, or the direction of the pathway changed through an angle of more than 60°.

Following whole-brain tractography, the fornix was extracted by drawing several regions of interest (ROIs) defined according to previously published methods (Metzler-Baddeley et al., 2011, 2013). The ROIs were drawn manually for one subject and then applied to all of the other subjects using an "atlas-based" tractography (ABT) approach, by spatial transformation of the ROIs to the other subjects' native space. This ensured consistency in the placement of the ROIs. See **Figure 1** for an example of the ROIs in the template subject, which was a HC participant. For several subjects this ABT approach was unsuccessful, most often due to inter-subject anatomical variability, such as neurodegeneration and encroachment of the lateral ventricles in some of the aMCI participants. In these cases, the ROIs for the tractography were drawn manually, with some adjustment and/or extra ROIs typically needed.

#### **Statistical analysis of the tractography data**

From the reconstructions of the fornix WM microstructural indices were extracted for statistical analyses in IBM SPSS Statistics for Windows, Version 22.0 (Armonk, NY: IBM Corp). These included free water corrected FA, corrected MD and the corrected Westin measures of linear diffusion coefficient (CL), and planar diffusion coefficient (CP) (Westin et al., 2002). The Westin measures describe the geometrical shape of the diffusion tensor, with a high value of CL implying only one dominant fiber orientation within a voxel (Vos et al., 2012), and a high value of CP indicating the presence of crossing fiber configurations (Vos et al., 2012).

The diffusion tensor metrics in the left and right fornix were compared within the two groups using a series of paired *t*-tests and no statistically significant differences were found. The values in the left and right fornix were therefore averaged to give a single fornix measure for each subject. This also reduced the number of multiple comparisons in the statistical tests, reducing the possibility of Type I errors. The diffusion metrics were compared between the groups using four paired *t*-tests, which were corrected for multiple comparisons using a Bonferroni-corrected *p*-level of *p* < 0.0125.

#### **RESTING-STATE FUNCTIONAL CONNECTIVITY OF THE THALAMUS AND HIPPOCAMPUS**

The fMRI data were processed using the DPARSF V2.3 toolbox (Data Processing Assistant for Resting-State fMRI, Yan and Zang, 2010) 2 , which utilizes SPM8<sup>3</sup> for the pre-processing steps and the REST V1.8 toolbox (Song et al., 2011) for the restingstate analysis. Data pre-processing involved slice timing correction, realignment to correct for head motion, normalization to MNI152 space by T1-image unified segmentation, smoothing with a 4 mm full-width-at-half-maximum Gaussian kernel, and detrending and filtering (0.01–0.08 Hz). The normalized voxel size was 3 mm × 3 mm × 3 mm. Several nuisance covariates were regressed out, including six head motion parameters and signals from the WM and CSF.

The resting-state FC of the thalamus and hippocampus was measured using a ROI FC approach. The ROIs were created using probabilistic atlases included in FSL<sup>4</sup> and were defined in the MNI152 template space. The left and right hippocampus ROIs were based on the Harvard-Oxford subcortical atlas structures. For the thalamus ROIs, the Oxford thalamic connectivity atlas (Behrens et al., 2003a,b), a probabilistic atlas of seven sub-thalamic regions, segmented according to their WM connectivity to cortical areas, was used to delineate the temporal region of the thalamus. The temporal region was chosen because, anatomically, this region is most likely to be structurally connected to the cortex nearest the hippocampus, and so the temporal thalamic-hippocampus resting-state measure was devised to mirror as closely as possible the regions structurally connected by the fornix. See **Figure 2**. The

<sup>2</sup>http://rfmri.org/DPARSF <sup>3</sup>http://www.fil.ion.ucl.ac.uk/spm

<sup>4</sup>http://fsl.fmrib.ox.ac.uk/fsl/fsl4.0/fslview/atlas-descriptions.html

**FIGURE 1 | Placement of the regions of interests (ROIs) for the dissection of the fornix in ExploreDTI**. **(A)** The five ROIs used to segment the fornix **(B)** coronal view of the seed ROI **(C)** ROI used to split the tracts and isolate the left fornix. The blue ROI is a SEED/OR gate, the green ROI is an AND gate, whilst the red ROIs are NOT gates. These ROIs were drawn for one subject and once deemed to be accurate and robust they were applied to all other subjects using an "atlas-based" tract segmentation method. The atlas in this case was the FA image of this subject.

four subcortical ROIs were thresholded at a minimum probability level of 20%, binarized, and re-sampled to 3 mm × 3 mm × 3 mm.

The average fMRI time series from the four ROIs were extracted and correlated to determine the FC of the regions at rest. Fisher-*Z* score transformations of these values were exported to IBM SPSS v.22 for second-level statistical analysis. The FC Fischer *Z*-scores between the left thalamus and left hippocampus, and between the right thalamus and right hippocampus, were compared within groups, and there was found to be no statistically significant difference. An average thalamus–hippocampus FC measure was therefore calculated for each subject, which was correlated against the fornix WM measures. For each group, four correlations were run and a Bonferroni-corrected *p*-value of 0.0125 was applied.

Several of the correlation coefficients were compared between the groups. To do this, the Pearson *r* values were first converted to Fisher *Z* scores. The *Z* statistic was then calculated using:

*<sup>Z</sup>* <sup>=</sup> (Fisher *<sup>Z</sup>*<sup>1</sup> <sup>−</sup> Fisher *<sup>Z</sup>*2)/<sup>√</sup> (variance), where variance is equal to: <sup>1</sup> *N*1 + 1 *N*2 .

#### **RESULTS**

#### **COMPARISON OF FORNIX WHITE MATTER MICROSTRUCTURE IN HC AND aMCI GROUPS**

The analyses of WM microstructural indices in the fornix revealed no statistically significant differences between the HC and aMCI groups. There was a trend towards higher CL and lower CP in the HCs versus aMCIs, however these results did not meet the Bonferroni-corrected threshold of *p* < 0.0125. See **Table 2** for a summary of these results and **Figure 3** for examples of the fornix tracts in one HC and one aMCI participant.

#### **RESTING-STATE FUNCTIONAL CONNECTIVITY OF THE THALAMUS AND HIPPOCAMPUS AND RELATION TO FORNIX WM MICROSTRUCTURE**

The average FC of the temporal thalamic regions and the hippocampus indicated a high degree of resting-state FC in both the HCs (Fisher's *Z* = 0.48 ± 0.23) and aMCIs (Fisher's *Z* = 0.50 ± 0.21). There was no statistically significant difference in FC between the groups [*t* (38) = −0.32, *p* = 0.75].

The correlational analysis of the functional and structural connectivity measures revealed a significant positive association in the


FA, fractional anisotropy; MD, mean diffusivity; CL, linear diffusion coefficient; CP, planar diffusion coefficient.

HC group between the linear diffusion coefficient in the fornix and the FC of the thalamus–hippocampus (*r* = 0.55, *p* = 0.008), which was absent in the aMCI group (*r* = −0.08,*p* = 0.81). These correlation coefficients differed significantly from one another (*Z* = 5.17, *p* < 0.0001). Scatterplots of these correlational results are shown in **Figure 4**. One of the datasets had a CL value of <0.2, which upon examination was found to be a statistical outlier (*p* < 0.05). This dataset was removed from the correlational analyses, thus in **Figure 4** there are 22 HCs and 17 aMCI datasets plotted.

#### **DISCUSSION**

In this study, we examined the structural integrity of the fornix, an important limbic WM tract, and the resting-state FC of two associated subcortical structures, the thalamus, and hippocampus. This is the first study to our knowledge to examine the relationship between the structural and FC of the thalamus and hippocampus, and whether this is altered in aMCI, a condition which is known to affect fornix WM and resting-state connectivity of intrinsic brain networks.

#### **RELATIONSHIP BETWEEN FUNCTIONAL AND STRUCTURAL CONNECTIVITY**

Although, contrary to our predictions, we found no evidence of reduced FC between the thalamus and hippocampus in aMCI, we did find that the correspondence between the FC and fornix WM was altered in this group. There was a significant positive relationship in the HCs between CL in the fornix and the FC of the thalamus–hippocampus, however this relationship was absent in the aMCIs.

Elucidating the relationship between structural and FC is not trivial, and is an area of growing interest within the neuroimaging community (van den Heuvel and Sporns, 2013; Sporns, 2014). Previous combined fMRI–DTI studies have found that resting-state FC does in general reflect the brain's structurally connected WM networks (Damoiseaux and Greicius, 2009; Greicius et al., 2009), and functional and structural connectivity measures have been found to correlate, particularly in healthy individuals (Skudlarski et al., 2008; van den Heuvel et al., 2008).

Several studies have indicated however, that the relationship between functional and structural connectivity can be altered in disease states. For example, in a study of major depressive disorder (de Kwaasteniet et al., 2013), patients had a negative correlation between FA in the uncinate fasciculus – another of the limbic WM tracts – and the FC between the subgenual anterior cingulate cortex and the hippocampus, which was not mirrored in the HC. Furthermore, in a study of schizophrenia (Skudlarski et al., 2010), patients

**FIGURE 3 | Example fornix tracts in a (A) healthy control and (B) amnestic aMCI subject**. The top panel in each case shows the tracts color-encoded with the first eigenvector (FE); the bottom panel shows the same tracts color-encoded with FA values. For each subject, the left and right fornix is shown in isolation on the left, and on the right the fornix can be seen overlaid on the subject's T1-weighted structural image, with the

FE shown in semi-transparent color (note the corpus callosum in red for example). The DWI data were co-registered to the structural images during the processing to correct for EPI-induced geometric distortions, however as illustrated by these figures this also facilitates the inspection of the white tracts in reference to the high-resolution anatomical image of the brain.

compared to controls were found to show not only overall reduced global WM connectivity as derived from DTI, but also a lower coherence between WM connectivity and FC. The current study supports the idea that the correspondence between functional and structural connectivity may be altered in neuropathological states. This is an area of research which certainly warrants further investigation in future studies.

Functional connections in the brain also exist between regions, which are not directly structurally connected, likely mediated by indirect structural connections (Honey et al., 2009), however in the current study we focused our analysis on the connectivity of two subcortical structures, which are known to be directly linked *via* a WM tract. We found evidence that there is a close correspondence between functional and structural connectivity in these structures in healthy ageing, which appears to be disrupted in aMCI.

#### **FORNIX RESULTS**

A number of previous studies have found that the fornix is affected in aMCI, with changes such as lower FA and increased RD reported (Lee et al., 2012; Metzler-Baddeley et al., 2012a; Oishi et al., 2012; Zhuang et al., 2013). We had predicted that the aMCI group would show similar changes in fornix WM indices consistent with compromised WM, however we found no evidence of this in the current study, with no differences in fornix WM microstructural indices between the groups.

The lack of difference in the fornix WM microstructural indices between the HCs and aMCIs may be due to sparing of the fornix in this aMCI cohort, however several other factors may also have contributed. This study was limited by a relatively small sample size, which likely reduced the power to detect group differences. Furthermore, the aMCI cohort included both single and multidomain aMCI. This increased heterogeneity of the sample may also have decreased the likelihood of detecting differences between the groups. Multi-domain aMCI also confers an increased risk of conversion to vascular dementia (VaD) (Petersen, 2004; Rasquin et al., 2005; Libon et al., 2010) and not just AD, which has typically not been found to involve the same level of disruption to the limbic system as AD (Burton et al., 2009). However some previous studies of aMCI also failed to detect changes in FA (Nowrangi et al., 2013; Rowley et al., 2013), suggesting that FA may not always be a strong indicator of WM connectivity changes in this cohort.

#### **LINEAR AND PLANAR DIFFUSION COEFFICIENTS**

The WM indices of linear and planar diffusion coefficients (CL and CP) did show a trend towards being significantly different between the groups (*p* < 0.035 in each case, which did not survive Bonferroni correction), with higher CL and lower CP in the HCs. High CL values are thought to reflect the presence of one dominant fiber orientation within a voxel (Vos et al., 2012), while high CP values are thought to indicate the presence of crossing fiber configurations (Vos et al., 2012). The fornix is generally considered to be a largely single-fiber orientation tract, however differences in CL and CP may also reflect WM damage. For example, in a recent study, which investigated WM changes following a season of playing varsity football (Davenport et al., 2014), there was a strong association between the number of head impacts and reduced FA, as well as reduced CL and increased CP. In fact, the strongest statistical relationship was with changes in CL, with the authors suggested may reflect WM disconnection due to the focal disruption of axons caused by the traumatic head impacts. Changes in CL and CP may also reflect WM changes in aMCI, such as WM degeneration and disruption to axonal organization, however further investigation of how these indices change in ageing and aMCI is needed.

#### **STRENGTHS AND LIMITATIONS OF THE CURRENT STUDY**

The present study extends previous studies by examining how fornix WM microstructural is related to the resting-state FC of two closely related structures, the thalamus and hippocampus. It is becoming common for neuroimaging researchers to collect multi-modal MRI data, therefore it is important for more studies to integrate findings across these modalities. This is the first study to our knowledge to combine FC and diffusion tractography to examine the fornix in ageing and aMCI, and the first to find differences in the relationship between the functional and structural connectivity of the thalamus and hippocampus related to aMCI.

The DWI methods employed in this study were very robust, as CSD-based tractography methods are preferable to DTI-based approaches, particular for brain regions with complex WM architecture. Correcting for free water is also an important preprocessing step since it is well established that atrophy-related CSF partial volume can bias DTI-based indices (Metzler-Baddeley et al., 2012a;Vos et al., 2012; Baron and Beaulieu, 2014;Maier-Hein et al., 2014). A recent study has highlighted how signals from CSF can contribute to differences in WM microstructural measures in MCI (Berlot et al., 2014), highlighting the importance of controlling for this confound.

A possible methodological limitation of our study is that we did not examine possible non-linear relationships between WM measures and FC. Variability in these measures may also be better captured by using more advanced statistical approaches, such as joint ICA (Calhoun et al., 2009). This is something which could be explored in future studies.

# **CONCLUSION**

In the current study, we used multi-modal MRI to examine the relationship between the functional and structural connectivity of two important subcortical structures, the thalamus and hippocampus, in healthy ageing and aMCI. We did not replicate previous findings of changes in DTI metrics in aMCI, which may indicate that the study was underpowered to properly test these differences. Thus the results should be taken as preliminary findings, which we believe warrant further investigation in a larger cohort. However, although there were no group differences in the WM measures of the fornix or in the FC of the thalamus– hippocampus at rest, there was a strong correspondence between the structural and FC measures in the HCs, which was absent in the aMCI group. The results suggest a disruption to the relationship between functional and structural connectivity in aMCI, which may be representative of early neuropathological connectivity changes in the limbic system. Both DWI and FC have offered new insights into brain network changes in aMCI. However, the complementary strengths of both of these methods combined may advance our understanding of neural network disconnection in this condition, and may increase the potential biomarker capabilities of MRI by elucidating subtle changes in the relationship between structural and functional brain networks.

# **ACKNOWLEDGMENTS**

We wish to acknowledge the assistance of Mr. Sojo Josephs in acquiring the MRI data. We would like to thank Prof. Alexander Leemans for advice regarding the DTI analysis, and Dr. Nia Goulden for advice on the use of the DPARSFA toolbox. The MRI data were accessed from the Lonsdale cluster maintained by the Trinity Centre for High Performance Computing. This cluster was funded through grants from Science Foundation Ireland. This research was supported by funding from a Department of Education and Learning N. Ireland (DELNI) Cross Border project between the University of Ulster and Trinity College Dublin, and from the European Regional Development Fund *via* the Interregional 4A Ireland Wales Programme 2007–2013.

# **REFERENCES**


Yan, C., and Zang,Y. (2010). DPARSF: a MATLAB toolbox for"pipeline"data analysis of resting-state fMRI. *Front. Syst. Neurosci.* 4, 13. doi:10.3389/fnsys.2010.00013 Yesavage, J. A. (1988). Geriatric depression scale. *Psychopharmacol. Bull.* 24, 709–711.


to episodic memory loss. *J. Alzheimers Dis.* 29, 629–639. doi:10.3233/jad-2012- 111766


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 31 July 2014; accepted: 22 January 2015; published online: 05 February 2015. Citation: Kehoe EG, Farrell D, Metzler-Baddeley C, Lawlor BA, Kenny RA, Lyons D, McNulty JP, Mullins PG, Coyle D and Bokde AL (2015) Fornix white matter is correlated with resting-state functional connectivity of the thalamus and hippocampus in healthy aging but not in mild cognitive impairment – a preliminary study. Front. Aging Neurosci. 7:10. doi: 10.3389/fnagi.2015.00010*

*This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2015 Kehoe, Farrell, Metzler-Baddeley, Lawlor, Kenny, Lyons, McNulty, Mullins, Coyle and Bokde. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Correlations between limbic white matter and cognitive function in temporal-lobe epilepsy, preliminary findings

#### **Ryan P. D. Alexander <sup>1</sup> , Luis Concha<sup>2</sup> ,Thomas J. Snyder <sup>3</sup> , Christian Beaulieu<sup>2</sup> and DonaldWilliam Gross <sup>1</sup>\***

<sup>1</sup> Division of Neurology, University of Alberta, Edmonton, AB, Canada

<sup>2</sup> Department of Biomedical Engineering, University of Alberta, Edmonton, AB, Canada

<sup>3</sup> Department of Psychiatry, University of Alberta, Edmonton, AB, Canada

#### **Edited by:**

Kenichi Oishi, Johns Hopkins University, USA

#### **Reviewed by:**

JackJ J. Lin, University of California, Irvine, USA Kenichi Oishi, Johns Hopkins University, USA

#### **\*Correspondence:**

Donald William Gross, Department of Medicine, Division of Neurology, 2E3.19 Walter C Mackenzie Health Sciences Centre, Edmonton, AB T6G 2B7, Canada e-mail: donald.gross@ualberta.ca

The limbic system is presumed to have a central role in cognitive performance, in particular memory. The purpose of this study was to investigate the relationship between limbic white matter microstructure and neuropsychological function in temporal-lobe epilepsy (TLE) patients using diffusion tensor imaging (DTI).Twenty-one adultTLE patients, including 7 non-lesional (nlTLE) and 14 with unilateral mesial temporal sclerosis (uTLE), were studied with both DTI and hippocampal T2 relaxometry. Correlations were performed between fractional anisotropy (FA) of the bilateral fornix and cingulum, hippocampalT2, neuropsychological tests. Positive correlations were observed in the whole group for the left fornix and processing speed index. In contrast, memory tests did not show significant correlations with DTI findings. Subgroup analysis demonstrated an association between the left fornix and processing speed in nlTLE but not uTLE. No correlations were observed between hippocampal T2 and test scores in either the TLE group as a whole or after subgroup analysis. Our findings suggest that integrity of the left fornix specifically is an important anatomical correlate of cognitive function in TLE patients, in particular patients with nlTLE.

**Keywords: mesial temporal sclerosis, temporal-lobe epilepsy, neuropsychological assessment, diffusion tensor imaging, processing speed**

#### **INTRODUCTION**

Temporal-lobe epilepsy (TLE) is the most common focal epilepsy syndrome and is often associated with cognitive comorbidity in particular for patients with mesial temporal sclerosis (MTS) (Hermann et al., 2009). Previous magnetic resonance imaging (MRI) volumetric studies have demonstrated that cognitive deficits in TLE are most strongly predicted by white matter volume (Hermann et al., 2002, 2003). Diffusion tensor imaging (DTI) permits the measurement of specific white matter tissue characteristics and thus has potential advantages over a non-specific technique like volumetric MRI (Basser et al., 1994). *In vivo* DTI of the fornix in TLE patients has shown histological correlates between fractional anisotropy (FA) and axonal membranes (Concha et al., 2010) thereby validating the technique as a non-invasive indicator of white matter micro-structural characteristics in human brain. Diffusion parameters of the uncinate fasciculus, inferior frontooccipital fasciculus, arcuate fasciculus, and cingulum have been demonstrated to correlate with verbal memory, naming performance, and fluency in TLE patients (Flugel et al., 2006; Diehl et al., 2008; McDonald et al., 2008; Riley et al., 2010). In the

**Abbreviations:** AVLT, Rey auditory verbal learning test recall score; CPMG, Carr Purcell Meibom Gill sequence; CVMT, continuous visual memory test total score; DTI, diffusion tensor imaging; FA, fractional anisotropy; MRI, magnetic resonance imaging; MTS, mesial temporal sclerosis; nlTLE, non-lesional temporal-lobe epilepsy; *r*(*p*) 0 , Spearman rho correlation.; ROI, region of interest; SD, standard deviation; TLE, temporal-lobe epilepsy; uTLE, temporal-lobe epilepsy with unilateral mesial temporal sclerosis; WAIS-III, Wechsler adult intelligence scale third edition.

present study, we focused on the fornix and cingulum, two prominent structures in the limbic white matter network, which have been demonstrated to be abnormal in TLE, albeit with differences between non-lesional TLE (nlTLE) and TLE with unilateral MTS (uTLE) (uTLE subjects have been demonstrated to have reduced FA of the fornix and cingulum while nlTLE have not) (Concha et al., 2005a,b, 2009). The purpose of this study was to assess correlations between white matter structure using DTI and cognitive function in a group of adult TLE patients.

#### **MATERIALS AND METHODS**

#### **STANDARD PROTOCOL APPROVALS, REGISTRATIONS, AND PATIENT CONSENTS**

Approval of the project was obtained from the University of Alberta Health Research Ethics Board and informed consent was obtained from all participants.

#### **SUBJECTS**

Twenty-one left hemisphere language dominant TLE patients were studied, including 7 nlTLE patients (3 left, 1 right, 3 bilateral) and 14 with uTLE (10 left, 4 right) (**Table 1**). No significant differences were observed between nlTLE and uTLE group means of age (nlTLE: 42, uTLE: 41, *p* = 0.86), onset age (nlTLE: 22, uTLE: 13, *p* = 0.10), or disease duration (nlTLE: 20, uTLE: 27, *p* = 0.21). The median number of prescribed anti-epileptic drugs per patient was two in both groups. Of note, a similar number of patients in each group were prescribed topiramate (nlTLE: 3, uTLE: 4), which has been shown to negatively impact cognitive ability, particularly verbal memory, and fluency (Thompson et al., 2000). TLE diagnosis



No significant differences were observed between nlTLE and uTLE groups for age (p = 0.83), education (p = 0.13) onset age (p = 0.09), and disease duration (p = 0.26).

was based on ictal semiology of complex partial seizures, confined temporal-lobe interictal, and ictal electroencephalogram findings. Absence of structural lesions outside of the temporal-lobe was confirmed on clinical MRI. MTS was defined based on hippocampal T2 greater than two standard deviations of controls. Patients were considered non-lesional if hippocampal T2 was within two standard deviations of control values. Left hemisphere language dominance was confirmed by either dichotic listening or intracarotid sodium amytal tests. The patient group used in this analysis is a subset of the study group previously reported (Concha et al., 2009) exclusions were made for: insufficient neuropsychological test data, neuropsychological testing greater than 36 months from the time of research MRI, or atypical language dominance. The median length of time between research MRI and neuropsychological evaluation was 1 month (range 0–35 months), 14 patients having less than 6 months between MRI and neuropsychological evaluation. Additional patient information has been previously reported (Concha et al., 2009).

#### **NEUROPSYCHOLOGICAL EVALUATION**

The neuropsychological tests used were: Processing Speed Index (Digit Symbol and Symbol Search subtests of the WAIS-III) (*n* = 19), that has been reported to be sensitive to white matter abnormalities (Axelrod et al., 2001; Drew et al., 2009), the Rey AuditoryVerbal Learning Test delayed recall score (AVLT) (*n* = 21) as a measure of verbal memory, and the Continuous Visual Memory Test total raw score (CVMT) as a measure of figural memory (*n* = 21). Standardized scores adjusted for both age and education were used.

#### **IMAGE ACQUISITION**

Cerebrospinal fluid-suppressed diffusion tensor images were acquired in 9:30 min using a 1.5 T Siemens Sonata MRI scanner (Siemens Medical Systems, Erlangen, Germany). The sequence consisted of 26 contiguous 2 mm thick axial slices positioned to provide coverage of the limbic tracts with an in-plane resolution of 2 mm × 2 mm (interpolated to 1 mm × 1 mm × 2 mm). Diffusion-sensitized images were acquired in six directions, with *a b* value of 1000 s/mm<sup>2</sup> . Full details of the DTI protocol have been previously provided (Concha et al., 2005a,b). T2 images for the quantification of hippocampal T2 were obtained using a modified CPMG sequence with 32 echoes (TR = 4.43 s; TE<sup>1</sup> = 9.1 ms, echo spacing = 9.1 ms), producing 10 coronal 3 mm thick slices with a 3 mm interslice gap in 8:13 min (voxel size 1.2 mm × 1.2 mm× 3 mm interpolated to 0.6 mm × 0.6 mm × 3 mm).

#### **IMAGE PROCESSING**

Bilateral fornix and cingulum were investigated using deterministic tractography. By transferring DTI images to a personal computer running DTIstudio 2.5 (Johns Hopkins University, Baltimore, MD, USA), tracts were depicted using fiber assignment by continuous tracking algorithm. Placement of tract-selection regions of interest (ROI) for each fiber bundle was based on methods outlined previously (Mori et al., 1999; Wakana et al., 2004). The FA threshold was set at 0.3 for all tracts. The FA threshold of 0.3 is commonly used in the deterministic tractography literature and is the threshold that we have used in our previous tractography studies of the fornix and cingulum (Mori et al., 1999; Concha et al., 2005a,b, 2009). If more than one streamline for a tract passed through a voxel, repeated coordinates were discarded and the voxel was only included once in the analysis. Mean diffusion parameters for each white matter tract were calculated in each patient. For this study, the crus of the fornix and the temporal portion of the cingulum were analyzed between the axial levels of the mammillary bodies and the fusion of the crura of the fornices. The T2 signal decay was fitted to a mono-exponential curve by voxel across the multi-echo coronal images. T2 values for each hippocampus were calculated by averaging within ROIs manually drawn on two consecutive slices (Concha et al., 2005a,b).

#### **STATISTICAL ANALYSIS Statistical tests**

Analyses were performed using SPSS 17.0 (SPSS Inc., Chicago, IL, USA). Spearman correlations were used to evaluate the rela-

tionship between FA of each tract and neuropsychological test scores. With a total of 12 tests performed for the primary analysis, Bonferroni correction sets a significance level of *p* < 0.0042.

#### **Primary analysis**

The primary analysis was made studying the entire group (nlTLE, uTLE, right, left, and bilateral TLE combined) based on the hypothesis that correlations would be observed between white matter measures and neuropsychological tests and that the previously reported differences in white matter DTI measures and neuropsychological tests seen in nlTLE and uTLE and right and left TLE would provide an adequate range of values to demonstrate significant correlations.

#### **Secondary analysis**

Where significant correlations were observed in the preliminary analysis, subsequent analysis was performed looking at the nlTLE and uTLE groups and right and left TLE groups separately. Correlations between hippocampal T2 and neuropsychological test scores were also assessed based on the hypothesis that changes in either white matter diffusion parameters or cognitive test scores could be secondary to mesial temporal pathology.

## **RESULTS**

#### **PATIENT NEUROPSYCHOLOGICAL TEST SCORES**

After examining whole group summary statistics of neuropsychological test scores [Processing Speed: 80 ± 15 (55–112); AVLT: 32 ± 12 (9–57); CVMT: 23 ± 17 (−11–63)], differences between nlTLE and uTLE subgroups were also investigated (**Table 1**).When compared to the average processing speed standard score of 100 (SD 15), the nlTLE group mean of 88 is low average but still within normal limits, whereas the uTLE group score of 76 is mildly deficient. Considering AVLT, for which the mean standard score is 50 (SD 10), both groups are within normal limits for AVLT but nlTLE patients scored higher (54) than uTLE (42). Both groups were impaired (<first percentile) on the CVMT (nlTLE: 24, uTLE: 23) (**Table 1**).

#### **PRIMARY ANALYSIS (ENTIRE TLE GROUP)**

Higher FA values for the left fornix were associated with higher Processing Speed indices [*r*(*p*) <sup>0</sup> = 0.62, *p* = 0.004] (**Table 2**). No other correlations were observed between Processing Speed, AVLT, CVMT, and FA of the fornix and cingulum (**Table 2**).

# **SECONDARY ANALYSIS**

#### **nlTLE and uTLE groups**

After splitting the TLE patients into subgroups (nlTLE *n* = 7, uTLE *n* = 14), other trends became evident. The correlation between FA of the left fornix and Processing Speed was explained by a strong correlation in the nlTLE group [*r*(*p*) <sup>0</sup> = 0.90, *p* = 0.02], whereas the uTLE group showed no correlation [*r*(*p*) <sup>0</sup> = 0.40, *p* = 0.18] (**Figure 1**).

**Table 2 | Spearman rho correlations of neuropsychological test scores and fractional anisotropy (FA) of the fornix and cingulum in TLE patients**.


Displayed as 'Spearmans's r (p)'. \*denotes significance with Bonferoni correction (p < 0.0042).

## **Right and left TLE**

Analysis of right TLE (*n* = 5) and left TLE (*n* = 13) demonstrated a positive correlation between FA of the left fornix and processing speed in left TLE [*r*(*p*) <sup>0</sup> = 0.60, *p* = 0.04].

## **Correlations between hippocampal T2 relaxometry and neuropsychology**

Correlations were not observed between any of the neuropsychological test scores and left or right hippocampal T2 in the total patient group. As well, no correlation was seen between left hippocampal T2 and processing speed in the nlTLE group.

# **DISCUSSION**

Quantitative MRI studies have shown white matter volume correlates with verbal and figural memory (Hermann et al., 2002, 2003). Recent DTI studies have demonstrated reduced FA in cerebral white matter in multiple brain regions and specific white matter tracts (Concha et al., 2005a,b, 2009; Gross et al., 2006; Focke et al., 2008; Schoene-Bake et al., 2009) with several reports demonstrating correlations between memory and DTI changes in TLE (Diehl et al., 2008; McDonald et al., 2008; Riley et al., 2010). While memory deficits are considered central to the clinical phenotype of TLE, broader cognitive measures such as intelligence, executive function and motor speed have also been demonstrated to be reduced (Hermann et al., 1997; Oyegbile et al., 2004).

The goal of this study was to investigate the relationship between limbic white matter microstructure and cognition in TLE. The primary observation of this study was that correlations were only observed between FA of the left fornix and processing speed for the TLE group as a whole with no correlations observed for the verbal (AVLT) and non-verbal (CVMT) memory tests nor for FA of the other tracts (right fornix and bilateral cingulum) and any of the cognitive measures.

It was observed that the Processing Speed Index, a WAIS-III measure that is most sensitive to brain disorders affecting white matter (Hawkins, 1998) such as traumatic brain injury (Axelrod et al., 2001) and multiple sclerosis (Drew et al., 2009), only correlated with left fornix FA whereas it is usually considered a measure of more widespread white matter function. Wernicke– Korsakoff syndrome, largely based in mammillary body pathology (Zubaran et al., 1997), has also been linked to deficits in one of the two subtests making up the Processing Speed Index, the Digit Symbol-Coding test, in addition to the expected deficit in memory test scores (Jacobson et al., 1990; Oscar-Berman et al., 2004). While there are no other specific references to the relationship between the left fornix specifically and Processing Speed, the link between Processing Speed and limbic structures both for Wernicke–Korsakoff as well as this study suggests an important role of the limbic system specifically in what has been considered a measure of more widespread white matter functioning. While the absence of correlations with other white matter structures suggests a unique role for the left fornix, given the limited sample size, false negative results cannot be ruled out, therefore,further work is required to better understand the relationship between Processing Speed and specific white matter tracts.

**test scores [Processing Speed Index, Auditory Verbal Learning Test (AVLT), and Continuous Visual Memory Test (CVMT)] and fractional anisotropy (FA) of the left fornix (A–C), right fornix (D–F), left cingulum (G–I), and right cingulum (J–L) in patients with temporal-lobe epilepsy** **sclerosis (uTLE)**. \*A significant positive correlation is seen between Processing Speed and FA of the left fornix which is primarily driven by the nlTLE subjects. No other significant correlations were observed for the other comparisons.

While caution must be taken in drawing strong conclusions from the secondary analysis, in particular given the small sample size of some of the subgroups, several interesting findings are observed in particular when looking at the nlTLE and uTLE subgroups.

We observed that the association between left fornix FA and Processing Speed is unique to the nlTLE group despite the greater deficit in intelligence scores previously shown to be evident in uTLE patients (McMillan et al., 1987). Of note, we have previously reported correlations between FA of the fornix and disease duration in nlTLE (Concha et al., 2009). Together these observations, albeit preliminary, are consistent with fornix degeneration and progressive cognitive dysfunction being secondary to ongoing seizures in nlTLE. Of note, the absence of correlations between hippocampal T2 and cognition suggests that the observed correlations between the left fornix and Processing Speed are not a secondary effect (i.e., our findings are not consistent with the nlTLE subjects developing hippocampal degeneration which then leads to downstream degeneration of the fornix and subsequent to this reduced FA).

The dissociation of findings between subgroups suggests the potential for seizure-related white matter and cognitive deterioration being unique to the nlTLE group, supporting the distinctiveness of the nlTLE and uTLE disease states. This idea has also been suggested and supported in recent DTI research (Concha et al., 2009; Kim et al., 2010; Shon et al., 2010). The fact that no correlations were observed in uTLE patients despite greater neuropsychological impairment suggests factors other than fornix damage are responsible for cognitive deficits in this group.

There has been much investigation into the extent and severity of cognitive decline in chronic TLE. Verbal memory peaks earlier and declines faster in TLE patients, especially in left TLE (Helmstaedter and Elger, 2009). A recent prospective study has demonstrated abnormal 4 year trajectory for memory as well as executive function and motor speed in TLE which was associated with baseline MRI abnormalities, lower baseline intelligence, older age, and longer duration of epilepsy (Hermann et al., 2006). As well, general intelligence has been demonstrated to be significantly reduced in earlier onset TLE, including patients without MTS (Kaaden and Helmstaedter, 2009). The cross-sectional nature of our data makes it difficult to draw conclusions regarding whether structural (fornix) or functional (processing speed) changes are progressive. Longitudinal studies are required to evaluate the relationship between functional and structural changes seen in TLE and determine whether nlTLE and uTLE may follow different trajectories.

With respect to division of our TLE sample, an appealing avenue of analysis would be to study group differences between right and left epileptic foci, as has been previously reported (Diehl et al., 2008; McDonald et al., 2008). The primary objective of this study was to use a population of patients with TLE to look for correlations between limbic white matter structure and cognitive function. The study design intentionally included uTLE, nlTLE as well as TLE patients with both right and left epileptic foci as this design was expected to provide a range of both structural and functional abnormalities [i.e., based on the literature uTLE patients were expected to have reduced limbic white matter FA (Concha et al., 2009) and lower cognitive test scores compared to nlTLE (McMillan et al., 1987) and right and left TLE patients were expected to expand the range of scores in particular tests of verbal and figural memory] (Powell et al., 2007). While it remains interesting to look at right and left differences, unfortunately, due to limitations of sample size in particular for the right TLE group (*n* = 5), it is difficult to compare right and left TLE in this study. While the left TLE group did show a significant correlation between FA of the left fornix and processing speed, the absence of positive findings in the right TLE group is most likely explained by the low sample size.

# **CONCLUSION**

In conclusion, our findings suggest that integrity of the left fornix specifically is an important anatomical correlate of cognitive function, in particular processing speed, in TLE patients. Furthermore, the differences in correlations of limbic white matter FA versus cognitive test scores in the subgroup analysis suggest that uTLE and nlTLE are distinctly different clinical and anatomical entities.

#### **ACKNOWLEDGMENTS**

Operating support provided by the Canadian Institutes of Health Research (DWG, Christian Beaulieu) as well as salary support by Alberta Innovates – Health Solutions (Christian Beaulieu).

# **REFERENCES**


data: a preliminary approach to clinical utility. *Clin Neuropsychol* 12, 535–551. doi:10.1076/clin.12.4.535.7236


tests of prefrontal brain functioning. *Alcohol. Clin. Exp. Res.* 28, 667–675. doi:10.1097/01.ALC.0000122761.09179.B9


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 11 March 2014; accepted: 12 June 2014; published online: 30 June 2014. Citation: Alexander RPD, Concha L, Snyder TJ, Beaulieu C and Gross DW (2014) Correlations between limbic white matter and cognitive function in temporal-lobe epilepsy, preliminary findings. Front. Aging Neurosci. 6:142. doi: 10.3389/fnagi.2014.00142 This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2014 Alexander, Concha, Snyder, Beaulieu and Gross. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Fractional anisotropy of the fornix and hippocampal atrophy in Alzheimer's disease

# **Kejal Kantarci \***

Department of Radiology, Mayo Clinic, Rochester, MN, USA

#### **Edited by:**

Kenichi Oishi, Johns Hopkins University, USA

#### **Reviewed by:**

Roberta Brinton, University of Southern California, USA Hui Wang, Children's National Medical Center, USA

#### **\*Correspondence:**

Kejal Kantarci, Department of Radiology, Mayo Clinic, 200 First Street SW, Rochester, MN 55905, USA

e-mail: kantarci.kejal@mayo.edu

Decrease in the directionality of water diffusion measured with fractional anisotropy (FA) on diffusion tensor imaging has been linked to loss of myelin and axons in the white matter. Fornix FA is consistently decreased in patients with mild cognitive impairment (MCI) and Alzheimer's disease (AD). Furthermore, decreased fornix FA is one of the earliest MRI abnormalities observed in cognitively normal individuals who are at an increased risk for AD, such as in pre-symptomatic carriers of familial AD mutations and in pre-clinical AD. Reductions of FA at these early stages, which predicted the decline in memory function. Fornix carries the efferent projections from the CA1 and CA3 pyramidal neurons of the hippocampus and subiculum, connecting these structures to the septal nuclei, anterior thalamic nucleus, mammillary bodies, and medial hypothalamus. Fornix also carries the afferent cholinergic and GABAergic projections from the medial septal nuclei and the adjacent diagonal band back to the medial temporal lobe, interconnecting the core limbic structures. Because fornix carries the axons projecting from the hippocampus, integrity of the fornix is in-part linked to the integrity of the hippocampus. In keeping with that, fornix FA is reduced in subjects with hippocampal atrophy, correlating with memory function.The literature on FA reductions in the fornix in the clinical spectrum of AD from pre-symptomatic carriers of familial AD mutations to pre-clinical AD, MCI, and dementia stages is reviewed.

**Keywords: fornix, DTI, Alzheimer's disease, hippocampus, MRI imaging**

# **ANATOMY OF THE HIPPOCAMPUS AND FORNIX CONNECTIONS**

Hippocampal projections to the subcortical structures that form the fimbria are carried by the fornix bundle. The fornix has two major components that connect hippocampus and subcortical structures. These are the larger postcommissural component and the smaller precommissural component. A majority of the axonal projections in the fornix originate from the subiculum of the hippocampus and form the postcommissural fornix carrying subicular axons to the mammillary bodies, which then project to the anterior nucleus of the thalamus. A major axonal tract originating from the anterior thalamus projects to the posterior cingulate cortex, which connects back to hippocampus via the cingulum tract closing the "Papez Circuit" (Benarroch, 2006). In addition, the precommissural fornix caries axons originating from the CA1 and CA3 pyramidal neurons of the hippocampus to the lateral septal nucleus, which projects to the medial septum. The medial septum and the nucleus of the diagonal band send cholinergic and GABAergic projections back to the medial temporal lobe via the fornix.

Damage to fornix in monkeys produces deficits in learning about the places of objects and about the places where responses should be made. They are impaired in using information about their place in an environment. Furthermore, fornix lesions impair conditional left–right discrimination learning based on visual appearance of an object on either side (Rolls, 2014). Fornix transection in human beings has been reported to produce significant and persistent anterograde amnesia (D'Esposito et al., 1995).

Because majority of the axonal projections carried by the fornix originate from the subiculum and to a lesser extent from CA1 and CA3 pyramidal neurons of the hippocampus, neurodegenerative processes impacting these neuronal populations would also lead to degeneration of the axonal projections originating from these neurons. However, it is important to note that not all axons in the fornix originate from the hippocampus, as the cholinergic and GABAergic connections originating from the medial septum and the nucleus of the diagonal band project back to the medial temporal lobe via the fornix. Therefore, degeneration of neuronal populations in the hippocampus may not fully explain the axonal loss in the fornix.

# **NORMAL AGING AND THE HIPPOCAMPUS–FORNIX AXIS**

In monkeys, number of myelinated nerve fibers in the fornix decrease with aging by 15–34%, but the loss of myelinated nerve fibers in the fornix is not disproportionately high or low compared to the rest of the white matter. However, it is not clear whether the myelinated nerve fiber loss from the fornix with age is due to loss of the parent neurons from the hippocampus, or due to age-related alterations that affect myelin sheaths, and alter the conduction velocities along nerve fibers resulting in degeneration in long projecting myelinated nerve fiber tracts. This loss of nerve fibers lead to a reduced subcortical connectivity of the hippocampus and may in-part be responsible for cognitive decline associated with normal aging even when the hippocampal neurons are intact (Peters et al., 2010).

Aging is associated with alterations in hippocampal dendritic morphology in the absence of neuronal loss, therefore age-associated structural alterations in the fornix less likely to originate from hippocampal neuronal loss, and more likely be the consequence of demyelination and degeneration of fornix fibers with age (Hof and Morrison, 2004). While both hippocampal volumes (Jack et al., 1998), and white matter fractional anisotropy (FA) decreases during aging (Walhovd et al., 2010), a relationship between fornix FA and hippocampal volume remains even after controlling for age. Using a recursive regression procedure, to evaluate sequential relationships between the alterations of the hippocampus and fornix, Pelletier et al. showed that hippocampal atrophy in healthy aging could be mediated by a loss of fornix connections potentially through a retrograde process (Pelletier et al., 2013). Longitudinal change in hippocampal volumes and fornix FA over the adult lifespan may be useful in determining whether hippocampal volume loss or a decrease in fornix FA occur earlier during the aging process.

Sex differences in structural network connectivity across the life span have been observed in DTI studies, with women showing greater overall cortical connectivity and network organization and greater efficiency than men (Gong et al., 2009). However, it is unclear at this time whether there is a difference in structural connectivity across the hippocampus-fornix axis among men and women.

#### **HIPPOCAMPUS-FORNIX AXIS IN PRODROMAL AND PRE-CLINICAL AD**

There is converging evidence from longitudinal studies on clinically followed and autopsied cohorts that most people with the amnestic form of mild cognitive impairment (MCI) who progress to dementia in the future, develop Alzheimer's disease (AD) (Flicker et al., 1991; Bowen et al., 1997; Morris et al., 2001; Bennett et al., 2002; Meyer et al., 2002; Jicha et al., 2006; Petersen et al., 2006). In keeping with these findings, both hippocampal atrophy and a reduction in fornix FA are frequently observed in amnestic MCI (Liu et al., 2011; Zhuang et al., 2012, 2013; Zhang et al., 2013) and are associated with memory decline (Mielke et al., 2012) and the risk of progression to AD (Mielke et al., 2012; Douaud et al., 2013).

The relationship between the structural integrity of hippocampus and fornix has been demonstrated in MCI (Mielke et al., 2012; Zhuang et al., 2012). In a cohort of cognitively normal (CN) older adults (*n* = 570; median age = 78; interquartile range = 74-83) and MCI (*n* = 131; median age = 80; interquartile range = 77– 86) from the community, both CN and MCI patients showed decreased fornix FA if they also had hippocampal atrophy or AD pattern of hypometabolism on <sup>18</sup>F Fluorodeoxyglucose (FDG) PET compared to the CN group with normal hippocampal volumes and metabolism on FDG PET. Having only a positive amyloid-β PET scan with Pittsburgh compound-B (PiB) was not responsible for a reduction in fornix FA in CN individuals, demonstrating that high amyloid load does not influence diffusion tensor imaging (DTI)-based measures of white matter integrity in the absence of co-existent gray matter neurodegeneration in nondemented older adults (**Figure 1**) (Kantarci et al., 2014). Although individuals with MCI and pre-clinical AD have reductions in fornix FA, it should be noted that a reduction in fornix FA by itself would be insufficient in classifying preclinical and prodromal AD. However, fornix FA can be combined with other imaging

**FIGURE 1 | Fractional anisotropy of the fornix in the cognitively normal (CN) and mild cognitive impairment (MCI) biomarker groups**. Hippocampal atrophy on MRI and/or hypometabolism in the Alzheimer signature composite on FDG PET was used to classify subjects into the neurodegeneration-positive group, and high amyloid load on PET was used to classify subjects into the amyloid-positive group. Cut-points for amyloid positivity, hippocampal atrophy, and Alzheimer signature hypometabolism were determined from the 10th percentile of the

measurement distributions in clinically diagnosed AD patients as previously described (Jack et al., 2012). Fornix FA was lower in the neurodegeneration positive CN and MCI patients compared to the cognitively normal group with normal imaging findings (amyloid and neurodegeneration negative; p < 0.001). A positive amyloid PET scan was not associated with a reduction in fornix FA, in the absence of co-existent neurodegeneration in cognitively normal individuals (p = 0.19) (Kantarci et al., 2014).

biomarkers such as hypometabolism on FDG PET, hippocampal atrophy, and amyloid-β PET for a more accurate classification of individuals at risk for AD, and predicting outcomes.

In carriers of the fully penetrant familial AD mutations (Ringman et al., 2007), a reduction in fornix FA was present in CN individuals who were destined to develop AD regardless of the cross-sectional area of the fornix, suggesting that the FA reduction in the fornix may precede hippocampal atrophy and clinical symptoms in AD. It is not possible to make inferences on whether FA reductions precede hippocampal atrophy on MRI in the course of AD based on findings from cross-sectional studies. Longitudinal investigations with serial DTI and structural MRI in pre-clinical AD and MCI may clarify the sequence of DTI and volumetric MRI findings in AD.

#### **HIPPOCAMPUS-FORNIX AXIS IN ALZHEIMER'S DISEASE**

There is a topographic concordance between gray and white matter diffusivity changes in AD (Kantarci et al., 2010), which implies that the disruption in white matter tracts is closely associated with the cortical pathology. Late-myelinating fibers in the brain such as the corticocortical association and the limbic pathways that are more vulnerable to degeneration related to AD than the earlymyelinating fibers such as the corticospinal tracts, which do not show any DTI abnormalities in early AD (Stricker et al., 2009; Kantarci et al., 2010; Oishi et al., 2012).

Consistent with the findings in normal aging and MCI and pre-clinical AD, correlations between fornix FA and hippocampal atrophy has been observed in AD (Firbank et al., 2007; Pievani et al., 2010; Lee et al., 2012). In a sub-regional analysis of the hippocampus Lee et al. reported that the strongest associations between the structural changes in the hippocampus and fornix diffusivity is localized to the hippocampal CA1 and anteromedial subiculum (Lee et al., 2012). CA1 and subiculum are the subregions of the hippocampus that are vulnerable to AD-related neurodegeneration (Davies et al., 1992). Hippocampal efferent fibers that pass through the fornix are thought to originate from both of these subfields. The authors concluded that:"Cortical neuronal damage and subcortical axonal defects in AD are likely to be closely linked with each other, possibly reflecting a suggested pathogenic interaction between the two" (Lee et al., 2012). However, they also acknowledged that this hypothesis needs to be tested through longitudinal studies.

#### **CONCLUSION**

The association of microstructural alterations in the hippocampus and fornix continues throughout human lifespan and remains even when these structures are impacted by AD starting from the preclinical and prodromal stages. Prospective studies investigating the rates of structural alterations in these two structures may clarify the temporal sequence of the reductions in fornix FA and hippocampal volumes through preclinical, prodromal, and dementia stages of AD.

#### **ACKNOWLEDGMENTS**

Dr. Kantarci's research program is supported by the R01 AG40042, P50 AG16574/Project1, and P50 AG44170/Project 2

#### **REFERENCES**


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 11 August 2014; paper pending published: 05 October 2014; accepted: 27 October 2014; published online: 13 November 2014.*

*Citation: Kantarci K (2014) Fractional anisotropy of the fornix and hippocampal atrophy in Alzheimer's disease. Front. Aging Neurosci. 6:316. doi: 10.3389/fnagi.2014.00316 This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2014 Kantarci. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# **Milap A. Nowrangi \* and Paul B. Rosenberg**

Department of Psychiatry and Behavioral Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, USA

#### **Edited by:**

Fernanda Laezza, University of Texas Medical Branch, USA

#### **Reviewed by:**

Bogdan O. Popescu, Colentina Clinical Hospital, Romania Shin Murakami, Touro University California, USA

#### **\*Correspondence:**

Milap A. Nowrangi, Department of Psychiatry, Johns Hopkins Bayview Medical Center, 5300 Alpha Commons Drive, 4th Floor, Baltimore, MD 21224, USA e-mail: mnowran1@jhmi.edu

The fornix is an integral white matter bundle located in the medial diencephalon and is part of the limbic structures. It serves a vital role in memory functions and as such has become the subject of recent research emphasis in Alzheimer's disease (AD) and mild cognitive impairment (MCI). As the characteristic pathological processes of AD progress, structural and functional changes to the medial temporal lobes and other regions become evident years before clinical symptoms are present.Though gray matter atrophy has been the most studied, degradation of white matter structures especially the fornix may precede these and has become detectable with use of diffusion tensor imaging (DTI) and other complimentary imaging techniques. Recent research utilizing DTI measurement of the fornix has shown good discriminability of diagnostic groups, particularly early and preclinical, as well as predictive power for incident MCI and AD. Stimulating and modulating fornix function by the way of DBS has been an exciting new area as pharmacological therapeutics has been slow to develop.

**Keywords: Alzheimer's, MCI, fornix, DTI, DBS**

#### **INTRODUCTION**

Alzheimer's disease (AD) is the most common neurodegenerative condition in aging. AD is a growing public health problem that is projected to reach epidemic proportions if disease-modifying therapies are not found. The latest figures from the Alzheimer's Association indicate that there are an estimated 5.3 million Americans living with AD. By 2050, an estimated 11–16 million people are expected to be diagnosed in the United States alone (Thies et al., 2013). Establishing the diagnosis of Mild Cognitive Impairment (MCI) and AD has evolved over the last 25 years with the most recent iteration of the National Institute on Aging and Alzheimer's Association criteria (Mckhann et al., 1984, 2011). These criteria place a new emphasis on the use of biomarkers of AD pathophysiology whereas the original criteria were based solely on the clinical evaluation. Although the use of biomarkers to establish diagnosis or track progression is considered a step forward in the field, there is a need for continued development of complementary technologies utilizing biological, physical, cognitive, and behavioral substrates.

Broadly, the initial pathologic changes in AD have been shown to involve the medial temporal lobes with the accumulation of neurofibrillary tangles and senile beta amyloid plaques (Bancher et al., 1993; Braak and Braak, 1995; Xu et al., 2000; Braak et al., 2006). This region is thought to mediate the retrieval and learning of semantic and episodic memory – the most common early cognitive deficits in AD (Aggleton and Brown, 1999; Behl et al., 2005; Levy and Chelune, 2007; Baldwin and Farias, 2009). The accumulation of tangles and plaques is associated with progressive atrophy of the cortical gray matter as loss of large pyramidal neurons (layers III and IV) advances, particularly in cortical association regions (Braak and Braak, 1995). In addition to this, several white matter (WM) abnormalities have been described and are thought

to reflect axonal disintegration, rarefaction, oligodendrocytosis, and astrocytosis (Xu et al., 2001; Roher et al., 2002; Shahani et al., 2002). Neuroimaging techniques, particularly magnetic resonance imaging (MRI), is a robust method used to visualize and detect subtle changes in structure and function of the substructures within the medial temporal lobe and related regions as important and perhaps early markers of disease and progression.

Though most volumetric imaging studies focused on atrophy of the entorhinal cortex and hippocampus, close inspection of the limbic structures has revealed significant volume reductions in patients with sporadic and familial AD (Decarli, 2001; Cash et al., 2013). The fornix is one structure within the limbic system that is receiving increasing attention recently in part because of its ease of detectability using MRI scanning as well as robust associations to cognitive changes, diagnostic group discrimination, and susceptibility to therapeutic intervention (Aggleton and Brown, 1999; Thomas et al., 2011). Because of the increasing interest in this structure, this brief review will serve as an updated survey of the key research over the last 10 years involving the fornix with hopes for continued and increasing efforts to better understand the use of this structure in the diagnosis, progression, and treatment of AD.

## **FORNIX – STRUCTURE AND FUNCTION**

The fornix is a WM bundle belonging to the medial diencephalon, which also includes the hippocampus, mammillary bodies, and anterior and medial thalamus. Grossly, the fornix is found on the medial aspects of the cerebral hemispheres connecting the medial temporal lobes to the hypothalamus. It is formed at first from the output fibers of the hippocampus in the mesial temporal lobe, beneath the floor of the lateral ventricle. It courses along the curve of the corpus callosum forming its body. At the level of the foramen of Monro, the fornix divides and travels inferiorly and posteriorly

in the lateral wall of the third ventricle to end in the hypothalamus and basal forebrain. Though the left and right fornices travel separately along their course from the hippocampus to the hypothalamus, they merge at the level of the forniceal body to form an important commissural tract (Nolte, 1993; Aggleton and Brown, 1999; Brewer, 2003; Patestas Ma, 2006). See **Figures 1A,B**.

The fornix is but one structure within the"temporal lobe memory system" also termed the Papez circuit described by James Papez in 1937 (Papez, 1995). The fornix has been demonstrated as the link between the hippocampus, mammillary bodies of the hypothalamus, and the anterior thalamic nuclei (via the mammilothalamic tract) (Nolte, 1993; Yamada et al., 1998; Aggleton and Brown, 1999; Patestas Ma, 2006). Efferent projections from the hippocampus to the medial diencephalon, particularly the anterior thalamic nuclei, are thought to be essential for normal hippocampal activity and are directed by the fornix, which then forms the major cholinergic output tract for the hippocampus (Cassel et al., 1997; Aggleton and Brown, 1999). Some of the earliest research involving lesion studies of this system asserted that damage to this axis resulted in anterograde amnesia, which occurs as a result of faulty encoding of episodic information (Aggleton and Brown, 1999). Specifically, damage to the hippocampus, fornix, and other diencephalic structures result in the inability to form declarative (semantic and episodic) memories. Although selective damage of the fornices has been shown to result in anterograde amnesia, the damage to other structures within the Papez circuit result in

similar deficits emphasizes that the integrity of the circuit rather than individual substructure is important.

#### **FORNIX AS AN IMAGING BIOMARKER IN MCI AND AD**

There has been increasing research in recent years on imaging the fornix as a biomarker for diagnosis, progression (or conversion), and cognition in MCI and AD. Some of the earliest research focused on fornix ultrastructure as a sign of MCI and AD. Much of this research employed MRI volumetric methods and showed atrophy (volume reduction) in many of the limbic lobe structures including the fornix. Early volumetric studies by Callen et al. (2001) showed that a number of structures within the Papez circuit (hippocampus, amygdala, anterior thalamus, hypothalamus, mammillary bodies, basal forebrain, septal area, fornix, orbitofrontal, and parahippocampal cortices) suffered significant atrophy with hippocampal and posterior cingulate regions being particularly affected. These early studies largely established complimentary findings to neuropathological observations of limbic atrophy in AD (Hopper and Vogel, 1976). The interest in establishing morphological changes in preclinical AD (MCI) as an early marker of disease was first undertaken by Copenhaver et al. (2006) who did not show significant atrophy of fornix or mammillary bodies in MCI as was initially hypothesized. The inability to show this difference was likely due to the relative insensitivity of volumetric methods used at the time even though imaging processes and analyses were improved upon compared to similar

tract; and 6. entorhinal cortex. Adapted from Nieuwenhuys (2008). **(C)** FA map

Oishi et al. (2011).

previous studies (Gale et al., 1993;Bilir et al., 1998;Kuzniecky et al., 1999; Callen et al., 2001). As a result, most recent research in the area has focused on fornix microstructure as the measurement of WM integrity. This renewed direction in imaging research of the fornix was based on the hypothesis that the earliest morphological changes occurred within medial temporal lobe gray matter, specifically the hippocampus and entorhinal structures with secondary effect to the efferent outflow tract through the fornix and mammillary bodies (Braak and Braak, 1995; Cassel et al., 1997; Aggleton and Brown, 1999). Needless to say, the underlying pathological processes affect an interconnected network of regions rather than single regions.

Diffusion tensor imaging has been the most commonly used *in vivo* MRI technique in studying the WM architecture of the fornix. Diffusion tensor imaging (DTI) is an indirect method of measuring WM integrity when conventional MRI lacks the contrast to delineate WM fiber tract organization (Beaulieu et al.,1996; Mori et al., 1999; Basser and Jones, 2002). Fractional anisotropy (FA) and mean diffusivity (MD) are two DTI measures used to quantify the integrity of WM microstructure by measuring the relative random motion of water in cerebral tissue. These measures are scalar values where higher values of FA and lower values of MD are thought to represent normal tissue cytoarchitecture where the random motion of water along a healthy axon, for example, is tightly constrained and restricted to movement along one direction (anisotropic). Two additional measures, axial diffusivity (DA) and radial diffusivity (DR) are associated with secondary degeneration of axons and breakdown of myelin and may provide additional detail into WM fiber integrity (Pierpaoli et al., 1996; Song et al., 2003, 2004).

Relating DTI measurement with other biomarkers that are considered more direct ways of detecting AD pathology has shown good correlation and has been a means of test validation. PET amyloid imaging is a well-established method of directly measuring AD pathology, β-amyloid (Aβ) deposition *in vivo,* but clinical application has been limited by convenience of use, invasiveness, and the half-life of radiotracers. When compared with DTI, several studies have shown positive correlation. A study by Chao et al. (2013) showed Aβ deposition as measured by Pittsburgh compound B (PiB) PET imaging was associated with reduced WM integrity (lower FA) in the fornix and splenium of the corpus callosum in subjects who ranged from normal to MCI. Similarly, Racine et al. (2014) showed a positive correlation between Aβ+ subjects and those with lower FA in the corpus callosum, hippocampal cingulum, and lateral fornix. Comparison with CSF markers of AD pathology Aβ<sup>42</sup> and p-Tau<sup>181</sup> has also been an important area of research. A study by Gold et al. (2014) showed a positive correlation between these CSF markers and WM integrity (FA) in the fornix, a relationship that persisted after controlling for hippocampal and fornix volume in a cohort of 20 cognitively normal older adults. Lower FA in the fornix was also associated with reduced performance on several cognitive tests including a Digit Symbol Test. This study is particularly important because CSF changes are thought to precede neuroimaging alterations by several years (Jack et al., 2010). These studies strongly suggest that microstructural measurements are related to AD pathology.

As the interest in the field of identifying early markers of AD continued to grow, several studies aimed to identify WM changes in preclinical AD. One of the earlier studies by Ringman et al. (2007) found decreased FA in the columns of the fornix were associated (linear regression) with hippocampal atrophy (*p* = 0.023), WM volume (*p* = 0.002), and mutation status (*p* = 0.032) in asymptomatic familial AD. They concluded that decreased FA in the forniceal columns is a robust finding in early familial AD and may be a biomarker of early disease in sporadic AD. Since this important study, there have been a number of more recent DTI studies showing similar associations between the WM tracts of the fornix and earlier preclinical stage. Cross-sectional DTI analysis by Mielke et al. (2009) and longitudinal analysis of the same cohort by Nowrangi et al. (2013) showed significant differences between normal participants and those with MCI and AD in the fornix, anterior cingulum, and splenium cross-sectionally and longitudinally (*p* < 0.01). Decreases in FA cross-sectionally and increases in MD longitudinally were observed in the fornix and were thought to indicate early markers of disease and longitudinal disease progression. Studies by Huang et al. (2012) and Boespflug et al. (2013) offered full DTI characterization (FA, MD, DA, and DR) of region of interest and atlas-based approaches to WM analysis respectively in MCI and AD. Both studies showed selective disruption of WM integrity in limbic structures including the fornix with DR being most sensitive to changes in the fornix. These studies further highlight a robust association between loss of integrity within the limbic structures and decrease in cognitive performance as illustrated by neuropsychological tests of episodic memory. Taken together, these studies suggest that the limbic WM tracts are preferentially affected in the early stages of cognitive dysfunction and that microstructural degradation of the fornix may precede atrophy of the hippocampus as detected by MRI.

Identifying preclinical signs of AD have been complemented by other research identifying individuals who would go on to progress to AD or convert from MCI to AD. One of the first studies observing this was by Oishi et al. (2011) who identified the fornix on an FA-scaled DTI map. FA reductions in this region predicted conversion form normal to amnestic MCI (aMCI) with a specificity of 1.0 and sensitivity of 0.67 and conversion from aMCI to AD with a specificity of 0.94 and sensitivity of 0.83. This finding was termed "the fornix sign" and has shown promise as an early predictive imaging sign of AD (See **Figures 1C,D**). Two other studies similarly emphasized the use of fornix microstructure as a predictor of conversion. One study byVan Bruggen et al. (2012)retrospectively used DTI analysis to differentiate between a subgroup of aMCI who converted to AD and a subgroup that did not. They found that FA, DR, and DA changes within the fornix, cingulum, and the corpus callosum were able to significantly discriminate between the two diagnostic groups and therefore could be thought of as a predictor for conversion between MCI and AD. Another study by Douaud et al. (2013) compared two groups of MCI patients; one that converted to AD no earlier than from baseline scanning and another that remained stable without progression of symptoms for at least 3 years. Comparing both volumetric and microstructural variables including DTI anisotropy and diffusivity, they found significant group differences in the body of the fornix, left fimbria, and the superior longitudinal fasciculus. These studies emphasize

the importance of microstructural integrity of the limbic lobe and specifically the fornix in predicting the progression/conversion from preclinical to clinical AD.

Recently, the field has been quite interested in identifying cognitively normal individuals who might be at increased risk for developing AD. The well-known and well-regarded hypothesis by Jack et al. and others that imaging changes are evident before clinical symptoms arise and that the pathological changes in the brain begin even decades before those has prompted increased interest in identifying these individuals (Morris et al., 1996; Bennett et al., 2006; Jack et al., 2010, 2013). In line with previous hypotheses that microstructural change precedes gross atrophic changes, there has been increasing work in using DTI and complimentary imaging methods to examine this. One study by Fletcher et al. (2013) identified the fornix as a region of microstructural change in cognitively normal elders. They found that the fornix body volume and DA were highly significant predictors of cognitive decline from normal cognition [*r* = =0.47 (0.24–0.89), *p* = 0.02, and *r* = =1.25 (1.02–1.46), *p* = 0.005, respectively]. This is one of the first studies establishing fornix degeneration as a predictor of incipient cognitive decline among healthy elderly individuals. Two studies conducted by Zhuang et al. (2012, 2013) demonstrated that WM changes in the fornix were evident in preclinical AD. One study (Zhuang et al., 2013) showed that when compared to


**Table 1 | Summary of key biomarker research of the fornix in cognitively normal (NC), MCI, and AD**.

cognitively normal subjects and those with late MCI, early aMCI subjects had lower WM integrity in the fornix. In a second study (Zhuang et al., 2012), the same group showed that when compared to normal subjects who remained stable over 2 years, those who converted to aMCI had substantial reductions in WM integrity in the fornix, parahippocampal gyrus WM, and cingulum while gray matter structures remained relatively intact. FA was found to be a predictor of conversion from cognitively normal to aMCI. Taken together, these studies emphasize the continued importance of detecting earlier preclinical forms of AD even in normal subjects in order to identify patients who might benefit from interventional approaches. **Table 1** summarizes the key studies described above.

## **FORNIX AND CLINICAL TREATMENT IN AD**

One of the most recent and exciting avenues of research has been directed at enhancing cognition by stimulating or modulating fornix function. Deep Brain Stimulation (DBS) has been thought of for some time now as having significant effect on mood and anxiety disorders such as major depression, obsessive compulsive disorder as well as Gilles de Tourette's syndrome, and other conditions. Similarly, it has been hypothesized that DBS may also play a role in enhancing memory functions. One of the first studies in this area by Hamani et al. (2008) used DBS to stimulate the fornix/hypothalamus area in a man with morbid obesity. In a case observation report, they also described that the procedure generated detailed autobiographical memory. In AD, this same concept has been hypothesized as being a potential therapeutic intervention for the treatment of amnesia and has just started being seriously studied. The first study of its kind, by Laxton et al. (2010), performed DBS stimulation of the fornix and hypothalamus in six AD patients. Though this was largely a safety study, the authors found that DBS triggered neural activity (increased metabolism) in some patients as seen through PET imaging (Laxton et al., 2010).

Later *in vivo* studies of DBS in patients with AD have been few and largely with relatively low *N*. Since 1985, there have been just two DBS studies targeting the stimulation of the fornix in patients with AD (Laxton et al., 2010; Fontaine et al., 2013). Found to over-all be safe, the clinical response has been rather modest with MMSE scores stabilizing rather than improving. However, reversal of AD-related hypometabolism has been observed in temporoparietal-occipital regions as well as increased connectivity between similar regions representing distributed cognitive networks (Smith et al., 2012). Given the lack of success thus far in developing pharmacological therapies for AD and the slow pace of further developments, such novel treatments as DBS is a promising avenue for therapy in AD. Currently, the Advance DBS clinical trial has recently completed recruitment of 42 participants with mild AD. Half of the study participants will have the stimulators turned on at the start of the study and the other half 12 months later. All participants will be followed for 18 months to track progression of their cognitive impairment. More information of this study can be found at www.clinicaltrials.gov.

#### **CONCLUSION**

There is a small but growing body of research directed toward the structure and function of the fornix in MCI and AD. Because of its important role within the "temporal lobe memory system," understanding the microstructural integrity of the fornix in MCI and AD has become an important area of focus in the field as there is a growing priority for identifying individuals early in the disease process and those who are at increased risk for progression or conversion. Recent studies have shown DTI is able to indirectly measure microstructural changes of the fornix that may precede gross atrophic changes of gray matter structures such as the hippocampus and entorhinal cortex. These studies demonstrate that DTI measurements of the fornix can significantly discriminate between cognitively normal, MCI, and AD groups thereby serving as a robust biomarker of disease. Moreover, fornix measurements have been shown to be a reliable measure of conversion from preclinical disease to AD. As pharmacological interventions continue to develop, studies of stimulating the fornix through DBS to alleviate or even reverse cognitive impairment has shown some early promise in stabilizing cognitive decline and increasing neural network activity.

# **REFERENCES**


workgroups on diagnostic guidelines for Alzheimer's disease.*Alzheimers Dement.* 7, 263–269. doi:10.1016/j.jalz.2011.03.005


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 September 2014; accepted: 02 January 2015; published online: 21 January 2015.*

*Citation: Nowrangi MA and Rosenberg PB (2015) The fornix in mild cognitive impairment and Alzheimer's disease. Front. Aging Neurosci. 7:1. doi: 10.3389/fnagi.2015.00001*

*This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2015 Nowrangi and Rosenberg . This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Early brain loss in circuits affected by Alzheimer's disease is predicted by fornix microstructure but may be independent of gray matter

#### **Evan Fletcher <sup>1</sup>\*, Owen Carmichael 1,2, Ofer Pasternak <sup>3</sup> , Klaus H. Maier-Hein<sup>4</sup> and Charles DeCarli <sup>1</sup>**

1 IDeA Laboratory, Department of Neurology, University of California Davis, Davis, CA, USA

<sup>2</sup> Department of Computer Science, University of California Davis, Davis, CA, USA

<sup>3</sup> Departments of Psychiatry and Radiology, Harvard University, Cambridge, MA, USA

<sup>4</sup> Medical Image Computing Group, German Cancer Research Center (DKFZ), Heidelberg, Germany

#### **Edited by:**

Kenichi Oishi, Johns Hopkins University, USA

# **Reviewed by:**

Claudia Metzler-Baddeley, Cardiff University, UK Michelle M. Mielke, Mayo Clinic, USA

#### **\*Correspondence:**

Evan Fletcher, IDeA Laboratory, Center for Neuroscience, University of California Davis, 1544 Newton Court, Davis, California 95616, USA e-mail: evanfletcher@gmail.com

In a cohort of community-recruited elderly subjects with normal cognition at initial evaluation, we found that baseline fornix white matter (WM) microstructure was significantly correlated with early volumetric longitudinal tissue change across a region of interest (called fornix significant ROI, fSROI), which overlaps circuits known to be selectively vulnerable to Alzheimer's dementia pathology. Other WM and gray matter regions had much weaker or non-existent associations with longitudinal tissue change. Tissue loss in fSROI was in turn a significant factor in a survival model of cognitive decline, as was baseline fornix microstructure. These findings suggest that WM deterioration in the fornix and tissue loss in fSROI may be the early beginnings of posterior limbic circuit and default mode network degeneration. We also found that gray matter baseline volumes in the entorhinal cortex and hippocampus predicted cognitive decline in survival models. But since GM regions did not also significantly predict brain-tissue loss, our results may imply a view in which early, prodromal deterioration appears as two quasi independent processes in white and gray matter regions of the limbic circuit crucial to memory.

**Keywords: fornix diffusivity, longitudinal brain change, limbic circuit, default mode network, normal cognition**

# **INTRODUCTION**

Recent work has shown that the fornix (Oishi et al., 2012; Fletcher et al., 2013a) and other memory-related white matter (WM) tracts (Zhuang et al., 2012, 2013) are sensitive predictors of conversion from normal cognition to mild cognitive impairment (MCI) or Alzheimer's dementia. In these findings, the loss of WM volume or microstructural integrity, in key structures, appears to be measurable before gray matter differences. Similarly, in groups of subjects with and without subjective memory complaints, but scoring normally on all cognitive tests – thus encompassing very subtle differences – WM diffusivity of the posterior cingulate, retrosplenial cortex (RSC), and precuneus (Selnes et al., 2012) and parahippocampal WM (Wang et al., 2012) was significantly different for those with memory complaints, while gray matter thickness was not. Recent results such as these have been taken to suggest that early WM deterioration may be at least partially separate (Agosta et al., 2011; Selnes et al., 2012) from the β-amyloid and tau pathologies of gray matter, whose evolution is known to characterize Alzheimer's disease (AD) (Braak and Braak, 1996), and have even raised the question whether Alzheimer's should be considered a disease of the WM (Sachdev et al., 2013).

On the other hand, the progression of β-amyloid and tau pathologies has long been established (Braak and Braak, 1996) and remains the dominant paradigm of AD modeling (Jack et al., 2010; Sperling et al., 2011; Jack and Holtzman, 2013). In these models, gray matter changes due to an "amyloid cascade" and the

tauopathy driven by it (Jack and Holtzman, 2013) underly brain atrophy and deterioration. The sequence of events consists of GM atrophy followed by WM disconnection and regional hypometabolism (Villain et al., 2008, 2010). This deterioration is best viewed as a network disconnection syndrome (Reid and Evans, 2013) in which specific networks – the Papez memory circuit (Villain et al., 2008; Acosta-Cabronero et al., 2010), also known as the limbic circuit (Nestor et al., 2003), together with the default mode network (DMN) (Greicius et al., 2004) – are selectively affected (Seeley et al., 2009). In recent biomarker models of AD progression (Jack and Holtzman, 2013), tauopathy may or may not arise before the amyloid cascade but increasing levels of β-amyloid, perhaps due to inadequate clearance from the brain, drive the trajectory of brain deterioration. Thus according to the "amyloid cascade hypothesis" (Jack and Holtzman, 2013), tauopathy is separate from amyloidopathy but dependent on it. When tauopathy progresses, it follows the well-known stereotypical pattern starting with the brainstem and transentorhinal cortex (Braak and Braak, 1996;Jack and Holtzman,2013). The biomarker models (Jack et al., 2010; Jack and Holtzman, 2013) therefore picture an ordering of amyloid and tau pathologies followed eventually by macroscopic brain changes and measurable declines in memory and clinical function.

Thus, there may be an apparent contrast between amyloid models and recent results, among preclinical subjects, suggesting the importance of WM differences as early markers of decline. This has raised the question whether the biomarker sequence of Jack and Holtzman (2013) should be modified to include measurable WM deterioration as one of the earliest events (Sachdev et al., 2013). It would therefore be useful to clarify the extent to which baseline white and gray matter measurements, in the prodromal phase preceding MCI and AD, are associated with longitudinal changes in the brain. This could help evaluate recent suggestions about the usefulness of key WM tracts as early biomarkers (Zhuang et al., 2012; Fletcher et al., 2013a; Sachdev et al., 2013) to supplement the CSF and brain biomarkers (Jack et al., 2010; Jack and Holtzman, 2013) of the amyloid models. And by comparing the sequencing of gray and WM changes, we may gain clues about whether white and gray matter processes are related or separate, thus advancing our knowledge of the ways in which the brain starts to change prior to measurable cognitive decline.

In this article, we investigate whether baseline measurements of cortical gray matter volume and WM microstructure, in regions of interests (ROIs) known to be associated with cognitive decline, are also predictors of early longitudinal brain changes in a cohort of clinically normal subjects. Our GM ROIs are the hippocampi, entorhinal cortical gray, and gray matter in the RSC and posterior cingulate – known sites of early and later AD pathology – as well as the anterior cingulate GM for comparison. WM ROIs include the fornix, which has been shown to be a predictor of earliest cognitive decline, along with the splenium and anterior and posterior cingulum, whose integrity is diminished in MCI or AD (Zhang et al., 2007; Mielke et al., 2009; Acosta-Cabronero et al., 2010). We also test the genu and corticospinal tract (CST) as WM tract controls. Our comparison of baseline WM with longitudinal change has some precedent. It is similar to the approach in an earlier study of middle-aged normal subjects (Ly et al., 2013), which found that baseline fractional anisotropy (FA) measurements in entorhinal WM, corpus callosum genu, and splenium were associated with tissue loss in the superior longitudinal fasciculus (SLF), corona radiata, and other regions. This indicated that baseline WM measurements may be predictive of WM tissue change even in relatively young subjects – 15 years before the lower age cutoff of our normal elderly cohort – and suggested that WM microstructure and tissue loss should be included in models of aging.

As in that study, we did not suppose an *a priori* hypothesis about *which* specific brain regions may be associated with baseline cortical GM thickness or microstructural features in the WM ROIs. But we did test the hypothesis that tissue loss in *some* brain area is associated with baseline measurements in one or more of these ROIs. And we analyzed these findings to attempt inferences about the nature and timing of brain processes preceding cognitive change in elderly normals.

## **MATERIALS AND METHODS PARTICIPANTS**

Subjects for this study consisted of cognitively normal individuals recruited into the Longitudinal Cohort of the Alzheimer's Disease Center of the University of California Davis (UCD ADC). Participants were recruited through community outreach using methods designed to enhance ethnic diversity (Hinton et al., 2010). All subjects provided informed consent before participating in this study.

#### **CLINICAL EVALUATION**

Each participant received multidisciplinary clinical evaluations at the UCD ADC at baseline and at approximately annual follow-up examinations. Evaluations included detailed medical history, with physical and neurological examinations. Diagnosis of cognitive syndromes – MCI or AD – was made according to standardized criteria (Mungas et al., 2011) by a consensus conference of clinicians. The clinical dementia rating sum of boxes score (CDRSum) (Morris, 1997) was assessed as a measure of clinically relevant functional impairment. Clinical "conversion" from normal cognition occurred when a subject with normal cognition at baseline was diagnosed as MCI or AD at a follow-up visit. The time to conversion was measured as the time from baseline to date of conversion.

#### **MAGNETIC RESONANCE IMAGING**

Each subject received at least two structural T1-weighted MRI scans, with the first one close to the date of the baseline clinical evaluation, as well as a diffusion MRI sequence at the same date as the first scan. All images were acquired at the University of California Davis, Imaging Research Center. The T1-weighted spoiled gradient recalled echo acquisition had the following parameters: TR 9.1 ms, flip angle 15°, field of view 24 cm, slice thickness 1.5 mm, and field strength 1.5 T. The 2D axial–oblique single-shot spin-echo planar imaging diffusion sequence had the following parameters: TE 9.4 ms, TR 8000 ms, flip angle 90°, field of view 24 cm, and slice thickness 5 mm. In-plane voxel dimensions were 1.875 mm in each direction. The *B* value was 1000 s/mm<sup>2</sup> with six gradient directions collected four times each (two times each in the plus and minus directions), plus two B0 images. In order to perform group statistical analysis, all subject T1 images at baseline were non-linearly deformed to a minimal deformation template (MDT) (Kochunov et al., 2001) using linear and non-linear Bspline transformation parameters (Rueckert et al.,2006). Diffusion images were corrected for eddy currents in native space using the FSL toolbox (Jenkinson et al., 2012). To account for the presence of extracellular free water contamination as partial voluming in the diffusion images, diffusion tensors were estimated by the free water elimination (FWE) algorithm (Pasternak et al., 2009). This algorithm estimates at each voxel the contributions to apparent diffusivity of separate tissue and free water compartments, returning both a voxel map of free water fractional volume (from 0 to 100%) and estimated diffusion tensors that represent tissue without free water. Scalar measures of FWE-corrected FA were calculated from the diffusion tensors. In addition to free water corrected maps, we previously calculated regular (uncorrected) FA maps that were warped to a template for group analysis. An FA–MDT was constructed as the average of 69 cognitively normal subject FA images, each linearly aligned to its corresponding native T1 image and then transformed to MDT space using the transformation parameters generated for its T1 image, after visual verification that the linear alignment between native FA and T1 was accurate. For the current study, native subject uncorrected FA baseline images were separately linearly aligned and B-spline deformed directly to the FA–MDT image that was already coincident with our MDT. The transformation parameters deforming native FA to FA–MDT were then used to deform individual subject

diffusivity images (FWE-corrected FA and Free water maps) into template space. In the remainder of this article, we refer to all FWE-corrected FA maps simply as FA.

#### **LONGITUDINAL CHANGE ANALYSIS**

Each normal subject received a T1-weighted structural scan at a date close to the baseline clinical evaluation and second scan at least a year later. Localized voxelwise longitudinal change between the two T1 images was computed using a tensor-based morphometry (TBM) algorithm designed to enhance sensitivity and specificity by incorporating knowledge of likely tissue boundary locations (Fletcher et al., 2013b). Brain change was quantified using log-Jacobians derived at each voxel from the TBM deformation field. In this analysis, the determinant of the 3 × 3 derivative matrix of the deformation field computes the local volume change factor indicated by the deformation. Determinants between 0 and 1 indicate volume shrinkage or loss, while determinants >1 indicate expansion. Thus, Jacobians represent multiplicative proportional volume changes from baseline. Performing a log transform then gives a distribution of values symmetric about 0, with negative logs indicating contraction, positive logs indicating expansion, and 0 indicating no volume change (Hua et al., 2008; Fletcher et al., 2013b). These will be referred to as the log-Jacobians. In longitudinal brain mapping of our elderly cohorts, log-Jacobians of CSF spaces like the ventricles and sulci are generally positive, while those located in tissue are negative, indicating patterns of tissue loss and CSF expansion. To perform group statistical analysis, native space log-Jacobian images were deformed into template space using the linear and B-spline parameters previously computed for native T1 images as described above. For valid group comparison, log-Jacobian images in template space were normalized to represent change over a 2-year period. This was done by multiplying each Jacobian image by a factor consisting of the ratio 2.0/∆T, where ∆T was the interscan time interval for that subject's two T1 MRI scans. This normalization is justified by the assumption, verified empirically on sample subjects having several scan times, that brain-tissue change is roughly log-linear over time.

#### **REGIONS OF INTEREST**

Regions of interest were used to evaluate the correlations of baseline measurements in both white and gray matter with longitudinal changes. For WM measurements, template MDT space ROIs were used to examine the association of mean FA with voxelwise brain changes. For the fornix body we used an ROI generated in-house by an experienced neurologist, and previously used in publications from our lab (Lee et al., 2012; Fletcher et al., 2013a). The genu and splenium of the corpus callosum and the anterior and posterior cingulum ROIs were generated by the same neurologist. We tested baseline mean FA values in these regions for the existence of significant clusters of associated tissue change as measured by our log-Jacobians. We also ascertained which relevant WM structures overlapped our clusters of significantly associated log-Jacobian change. We calculated overlaps of the Jacobian clusters with our thalamus, splenium, and cingulum ROIs, Brodmann areas (BA) also generated in-house, together with ROIs in the Johns Hopkins WM atlas (Zhang et al., 2010), all warped to our own MDT. For GM measurements, ROI GM volume was ascertained in each subject's native space. BA template MDT ROIs were inversely transformed to native space structural T1 images via a process we have used in previous publications (Fletcher et al., 2013a,b). GM volumes were computed from these ROIs overlaid over the native space four-tissue segmented image. Tissue segmentation in the native T1 images was accomplished using a Bayesian maximallikelihood algorithm enhanced for sensitivity at tissue boundaries (Fletcher et al., 2012). Hippocampal volumes were computed in native space by hand tracing methods that have been described previously (Carmichael et al., 2010; Lee et al., 2012). To test the hypothesis that baseline GM measurements were associated with longitudinal brain change, we used the hippocampi and BA ROIs that are known to be implicated in GM loss during the trajectory of cognitive decline – BA 28 and 34 (the ventral and dorsal bilateral entorhinal cortices).

#### **STATISTICAL ANALYSIS**

To assess the correlation of baseline ROI GM or WM measurements with longitudinal change, we used single linear regression models of ROI mean FA values or GM volumes and voxelwise log-Jacobian values, computing the Student's *t*-value at each voxel for the slope of the regression line. In these regressions, the slope of the regressions at brain-tissue voxels will generally be positive for ROI FA values, corresponding to the association between lower FA values (putative markers of impaired ROI microstructure) and more negative log-Jacobians (indicating more severe brain-tissue loss). Conversely, higher ROI FA values should correspond to less severe brain-tissue loss, computed by negative log-Jacobians of smaller magnitude. Similarly for GM ROIs, the associations should be positive in brain-tissue areas, corresponding to a smaller magnitude of local tissue loss associated with greater baseline GM volume. The *t*-value clusters were corrected for multiple comparisons (Nichols and Holmes, 2001). We did this by non-parametric permutation testing over 1000 iterations. We tested for size of contiguous voxel clusters all having regression *t*-value >3.5 for both the FA values and the GM volumes. Clusters with sizes in the top 95th percentile over all iterations were taken as regions of significant association (*p* < 0.05) between ROI FA or GM and longitudinal brain change.

To evaluate the significance of brain factors for cognitive conversion, we constructed Cox proportional hazards models (Cox, 1972). Each model contained demographic factors of age, gender, education, and ethnicity as covariates, together with a single brain measurement. A brain measurement was a significant factor in cognitive conversion if the probability of having a larger χ 2 than the one in the likelihood ratio test for that variable was <0.05.

# **RESULTS**

#### **SUBJECTS**

Longitudinal analysis was based on a group of 68 communityrecruited normal subjects with two structural MR scans, diffusion MRI taken at the baseline date, and clinical data gathered at or near the dates of the structural MRIs. This group contained 12 subjects who later converted to MCI or AD, and 56 who remained stable. Demographics of the subjects are displayed in **Table 1**. Except for information on clinical conversion, all other data in the table reflect baseline date evaluations.

#### **LONGITUDINAL BRAIN-TISSUE LOSS ASSOCIATED WITH BASELINE WM AND GM ROIs**

In the regression tests of ROI GM volume (including hippocampal volume) and voxelwise log-Jacobians, we found only one small significant cluster (2 cm<sup>3</sup> ) log-Jacobian cluster for *t* > 3.5, corresponding to hippocampal volume. Other GM ROIs had no significant clusters.

For the correlations with WM ROIs, **Figure 1** displays raw *t*values of log-Jacobians vs. FA in (left to right) the fornix body, splenium, and posterior cingulum. A consistent pattern of highly positive correlations is seen for the fornix (left panel) as compared to inconsistent and low magnitude associations for the other two ROIs in the middle and right panels. Sizes of significant clusters for all WM ROIs, corresponding to a threshold of *t* > 3.5, are given in **Table 2**. None of the anterior and posterior cingulate, the splenium, and genu ROIs had any significant clusters for *t*-values above this threshold. The CST had a very small cluster. The fornix body generated much larger ROI comprising components bilaterally in the posterior parietal, temporal, and frontal lobes, with small sections in the cerebellum. We will henceforth designate the ensemble of these clusters as *fornix significant ROI* (fSROI). Its

#### **Table 1 | Demographics of normal cohort having DTI**.


For ethnicity designations:W means white, H means Hispanic, AA means African American. Starred variables are significantly different between converters and non-converters.

tissue composition consisted of about 80% white and 20% gray matter. The fSROI is shown in **Figure 2**.

#### **ANATOMICAL COMPOSITION OF THE fSROI**

**Table 3** shows the intersection of the fSROI with selected BA, regions of the corpus callosum, thalamus, longitudinal fasciculi, and WM tracts from the CC and thalamus to the parietal and occipital lobes. It indicates extensive overlap with the posterior cingulate cortex (PCC) (BA 23 and 31) and RSC (BA 26–29 and 30), as well as the thalamus. More generally, the fSROI heavily intersects the posterior portions of the corpus callosum, longitudinal fasciculi, and several of the WM tracts from thalamus or CC to posterior parietal or occipital gyri. Color-coded maps of intersections with ROIs listed in **Table 3** are shown in **Figure 3**. Green areas illustrate extensive overlap of BA; purple shows overlap with thalamus and its projections; and tan shows coverage of the splenium and projections from the corpus callosum.

#### **BRAIN FACTORS AS PREDICTORS OF COGNITIVE CONVERSION**

We tested the significance of the fSROI along with baseline brain measurements in Cox proportional hazards models for cognitive conversion from normal. Each model contained the multiple demographic variables described previously along with a single brain factor. Significant factors in the individual models, with their *p*-values and those of significant covariates, are summarized in **Table 4**. To facilitate effect comparisons amongst all brain variables and with age, hazard ratios for each brain factor are provided based on unit change in *Z*-score for brain factors and 1 year increase for age. Significant GM predictors of conversion were volumes in the ventral and dorsal entorhinal regions (BA 28, 34), the dorsal anterior cingulate (BA 32), and bilateral hippocampal volume. Age and ethnicity were also significant factors in the GM models. Fornix body FA and mean tissue change in fSROI were the other significant brain predictors of conversion. We note that in the GM models, age was highly significant and except for BA 28, it had a lower *p*-value than the accompanying brain factor. In the models of fornix FA and fSROI, by contrast, age was not significant.

#### **INTERACTIONS BETWEEN GM, WM, AND fSROI TISSUE LOSS**

To test for confounding effects of age or other demographic factors in association between baseline fornix FA and fSROI log-Jacobians,

**FIGURE 1 | Axial views of raw t-values from voxel regressions of log-Jacobians vs. FA**. Left: regressions with fornix body FA. Middle: regressions with splenium. Right: regressions with posterior cingulate. Bright indicates positive t-values, reflecting correlation of high ROI FA with smaller brain tissue loss.

we modeled the mean fSROI values as outcome of a multiregression including independent variables of age, gender, education, ethnicity, and fornix FA. The overall model produced a strong fit (*R* <sup>2</sup> = 0.52) with significance of the independent variables as follows: fornix FA, *p* = 0.0008; age, *p* = 0.016; education, *p* = 0.17; gender, *p* = 0.23; and ethnicity *p* = 0.69. Thus, fornix FA and age were the only significant correlates of tissue change in fSROI, with FA being much stronger than age.

To test for GM contributions to fSROI outcomes, we performed multiple regression modeling of mean fSROI log-Jacobians vs. baseline hippocampal, BA 28 or BA 34 gray volumes, one per regression model, together with the demographic factors of age, education, gender, and ethnicity. In each of these models, the overall fit was not as good as for the fornix FA (*R* <sup>2</sup> was about 0.36 in each case) and no brain factor was significant, while age was significant in each model.

Finally, to test the mutual interaction of the brain variables, we examined the correlation matrix of components including the fSROI log-Jacobians, fornix FA, and GM volumes from BA 28, 34, and bilateral hippocampus. The matrix is given in **Table 5**.

# **DISCUSSION**

In this study, an analysis of the relation between baseline ROI FA measurements and longitudinal brain change has indicated that fornix body microstructure is significantly associated with a large

**Table 2 | Volumes of significant clusters in regression of ROI FA values with voxel log-Jacobians, for t** > **3.5**.


There were no significant clusters for any GM ROI corresponding to threshold t > 3.5 except for the hippocampi, with cluster volume 2 cm<sup>3</sup> .

region of tissue loss (fSROI) primarily in the WM of the posterior frontal, parietal, and temporal lobes. Similar correlation analyses of WM in the CST, anterior and posterior cingulate, genu, and splenium, showed small or non-existent effects. Baseline GM values for the entorhinal cortex and hippocampus also failed to show strong associations with longitudinal brain change. The fornix has already been identified as a sensitive predictor of early cognitive change in normals (Oishi et al., 2012; Zhuang et al., 2012; Fletcher et al., 2013a) – a result corroborated in our present study – and our current findings now also link it with longitudinal change in posterior brain structures known to be affected by AD pathology (**Table 3**; **Figure 3**). Our regressions incorporating age as a covariate have indicated that these associations are not merely dependent on age. And the tissue losses in the fSROI are themselves significant predictors of conversion, in Cox survival models that included age as a covariate. The picture is nuanced, however, because GM regions in the limbic circuit also predicted cognitive conversion (**Table 4**) but not longitudinal brain change. In this section, we discuss these findings in relation to previous work, and attempt to draw some inferences about the nature and timing of early brain changes associated with cognitive decline.

#### **fSROI, LIMBIC CIRCUIT, AND PRIOR RESULTS CHARACTERIZING EARLY AD**

**Table 3** and **Figure 3** outline the intersection of the fSROI with other brain regions. The fSROI has relatively large overlaps with the thalamus; projections of the thalamus to precuneus; the superior parietal gyrus (including precuneus) and underlying WM; the RSC (BA 26, 29, and 30); the posterior cingulum (BA 23 and 31); and the splenium of the corpus callosum. Most of these are components of the circuit of Papez or limbic circuit (Nestor et al., 2003), considered to be important in the formation and consolidation of memory (Thomas et al., 2011). The fSROI is similar to the ROI found in a previous study investigating disconnection of the hippocampus and PCC in a cohort of early AD subjects (Villain et al., 2008). Their ROI of correlations between WM and hippocampal GM *Z*-scores (*p* < 0.05 uncorrected) included most of the cingulum bundle from frontal to parahippocampal fibers. They reported a correlation between gray matter *Z*-scores of the hippocampus and damage to the retrosplenial BA 29, 30, together with WM disruption of the posterior cingulum bundle. Cingulum bundle

**FIGURE 2 | Significant cluster of association (fSROI) between fornix body FA mean and negative log-Jacobians (t** > **3.5)**. Views from left to right: **(A)** axial, **(B)** coronal, and **(C)** sagittal.

#### **Table 3 | Component brain regions, which overlap with the whole fornix significant association ROI (fSROI)**.




Right column shows what percentage of the given ROI is intersected by fSROI. (A) Brodmann areas intersecting fSROI. (B) Corpus callosum (CC) and tracts originating in CC. (C) Thalamus (TH) and tracts originating there. (D) Longitudinal fasciculi; CC\_SOG, white matter tract from CC to superior occipital gyrus, etc.; IFOF, inferior frontal occipital fasciculus; TH\_PrCu, thalamo-precuneus tract, etc.; ILF, inferior longitudinal fasciculus; SLF, superior longitudinal fasciculus; SPG, superior parietal gyrus (encompasses precuneus and connections from CC and thalamus); RSC, retrosplenial cortex and underlying white matter (encompasses Brodmann areas 26, 29, and 30).

ILF 22 SLF 34

atrophy was in turn found to be correlated with glucose hypometabolism of the thalamus, mammillary bodies, middle cingulum, parahippocampal cingulum, and hippocampus – again components of the Papez circuit. The difference from our study was that in their early AD subjects the ROI arose from correlations of hippocampal and WM disruption, whereas our similar fSROI was correlated with fornix microstructural integrity at baseline. We shall discuss this below.

Another study (Acosta-Cabronero et al., 2010) examined the degeneration of WM tracts in early AD, using axial and radial diffusivities and mean diffusivity (MD). It found degeneration of tracts in the Papez circuit including the fornix, splenium, posterior cingulum, parahippocampal gyrus, and adjacent tracts of the temporo-parietal cortex as compared with normal controls. Anterior WM tracts were relatively spared. An important conclusion, according to that paper, is that this limbic circuit is preferentially vulnerable to degeneration in AD. This echoes an earlier proposal, based on the study of hypometabolism in MCI and AD subjects, that damage to the limbic circuit is the earliest significant event in the progression of AD (Nestor et al., 2003).

To summarize, then, previous studies both of hypometabolism and WM tract disruption among MCI and AD cohorts have found degeneration in regions which our fSROI extensively overlaps and in fact resembles. This is consistent with the possibility that the fSROI represents early longitudinal tissue changes, predicted by baseline fornix FA, which precede and eventually result in the effects described in those studies. This possibility is strengthened by our finding that brain-tissue losses in the fSROI are themselves significantly associated with cognitive conversion in survival models (**Table 4**).

#### **fSROI AND THE DEFAULT MODE NETWORK**

**Table 3** and **Figure 3** show that the fSROI extensively overlaps the RSC. The RSC has been identified with the one of the main nodes, the PCC/RSC node (Greicius et al., 2009) of the DMN (Greicius et al., 2003), which is connected to a second principal node, the medial prefrontal cortex, by the cingulum bundles (Greicius et al., 2009). The RSC is also connected more ventrally to the medial temporal lobes probably by fibers in the descending (parahippocampal) cingulum (Thomas et al., 2011). In AD subjects, the DMN has diminished resting state activity in the PCC (Greicius et al., 2004), probably resulting from disrupted connectivity between the hippocampus and posterior cingulum, and DMN activity may be a sensitive biomarker for incipient AD. Further, among five distinct functional connectivity networks of the brain selectively targeted by differing neurodegenerative diseases, the DMN is the one specifically characterized by vulnerability to AD (Seeley et al., 2009). Because the fSROI extensively overlaps the RSC and posterior cingulum, it appears that the fSROI contains early tissue loss in principal posterior components of the DMN, thereby presaging the later loss of connectivity and decreased activity. And the strong association of the fSROI with fornix FA therefore suggests that fornix microstructure in cognitively normal subjects may be a very early marker of future disconnection in the DMN.

#### **INFERENCES ABOUT EARLY BRAIN PROCESSES OF DEGENERATION**

The fornix and the rest of the limbic circuit, along with the DMN, are all selectively vulnerable to early Alzheimer's pathology (Nestor et al., 2003; Seeley et al., 2009; Acosta-Cabronero et al., 2010). And degeneration in the limbic circuit has been proposed as the first significant event in the progression of AD (Nestor et al., 2003). If our identification of the fSROI with major components of the limbic circuit and DMN is correct, then we may have traced the trajectory of tissue losses in these circuits farther backwards toward their origins than has been done heretofore. In this early phase, the changes are associated with baseline fornix WM but




Each model contained age, gender, education, and ethnicity together with one brain factor. p-Values refer to the probability of the χ <sup>2</sup> value being larger than computed for a given variable in likelihood ratio test. All brain factor values were converted to Z-scores prior to analysis in order to enable meaningful comparisons of hazard ratios. The hazard ratio columns show change in conversion risk per unit Z-score change (in brain factors) or year (age). Values in parentheses for these columns are the lower and upper 95% confidence interval endpoints. GM regions BA 26–29, 23, 31 are not shown because not significant; nor were the FA white matter regions in anterior and posterior cingulate, genu, and splenium; BA, Brodmann area; AC, anterior cingulate; ER, entorhinal cortex; fSROI, fornix significant ROI.

#### **Table 5 | Correlation matrix for factor interaction analysis of GM, fornix FA, and fSROI variables**.


From the table, fSROI is strongly correlated (0.60) with fornix FA but with no GM factor or the hippocampus. But fornix FA, while being most strongly correlated with fSROI, is also correlated with BA 28 and 34 (0.38 and 0.45, respectively). And BA 28 and 34 are strongly correlated with each other, and less so with hippocampal volume.

not strongly with temporal gray matter. Tissue losses in fSROI are greater than due to aging alone, since their correlation with fornix FA was much stronger than with age. On the other hand, baseline GM measurements in the hippocampus and entorhinal cortices (also components of the limbic circuit) were significant

predictors of cognitive decline but not of any brain-tissue loss. Analysis of these components (**Table 5**) showed that entorhinal areas were correlated with each other and with fornix FA but not strongly with fSROI. This may support the inference of two processes in early deterioration. One process may contribute most heavily to fornix WM and fSROI changes, underlying that strong correlation. The other may contribute to GM in the temporal lobe ROIs and also to fornix WM, but little to fSROI. Thus early deterioration in the posterior limbic circuit may be at least partly separate from gray matter processes in the entorhinal cortex or hippocampi. This inference does not necessarily contradict previous findings relating fornix and hippocampus (Lee et al., 2012) in MCI and AD, or cingulate WM and hippocampus (Villain et al., 2008, 2010) in early AD, because in the later stages, gray matter atrophy is more advanced and may contribute to circuit deterioration. Furthermore, a previous study has suggested (Sexton et al., 2010) that though fornix and gray matter pathologies may not be independent, the relationship may be complex, involving factors of Wallerian degeneration as well as separate WM disruptive processes. And another study (Douaud et al., 2013) has found that more than 2 years prior to conversion, WM microstructural differences in the fornix, left fimbria, and SLF characterized MCI patients who converted to probable dementia, when compared

with other stable MCI subjects. Those findings appear to be consistent with the suggestions put forth here, that in the prodromal phase, fornix microstructure, and tissue loss in the limbic circuit overlain by fSROI may have a component independent of gray matter change.

#### **LIMITATIONS**

We must note several limitations of our study. The first limitation is the small size of our normal cohort. This may have restricted our statistical power, for example in seeing relationships between GM ROIs and longitudinal brain change. The second limitation is the low resolution of our diffusion MRI data and the uncertainty of using it to make inferences about a structure as small as the fornix. It is evident from recent studies (Metzler-Baddeley et al., 2012; Berlot et al., 2014) that caution is required when interpreting diffusivity measurements, because of free water contamination to WM tracts generally and particularly in the fornix, due to its small size and proximity to the ventricles. But FWE (Pasternak et al., 2009) seems to improve FA's distinctions between groups and the robustness of inferences about WM change (Metzler-Baddeley et al., 2012). In our experiments, FWE improved the statistical power of FA correlations with longitudinal brain change, enabling us to see a more clear contrast between the effects of the fornix and other WM tracts (**Figure 1**), while dampening diffusivities such as MD and their apparent statistical power (data not shown). Lastly, we were unable in this project to explore hypotheses about the possible origins of WM changes in the fornix and the associated tissue loss in the fSROI,and whether such changes are related to gray matter changes by some underlying mechanism. A likely candidate for such exploration would be brain levels of β-amyloid accumulation.

# **CONCLUSION**

In a cohort of normal subjects, baseline fornix WM microstructure significantly correlates with early longitudinal tissue change across an fSROI intersecting the posterior DMN and limbic circuits. These circuits are known to be selectively vulnerable to AD pathology. Baseline fornix WM microstructure, longitudinal tissue loss in the fSROI, and baseline GM volumes in the entorhinal cortices and hippocampus all are significantly associated with cognitive conversion from normal. But only fornix WM is associated with brain-tissue loss in the fSROI. Thus, WM deterioration in the fornix and tissue loss in fSROI may be the early beginnings of posterior limbic circuit and DMN degeneration. This early degeneration appears to be partly separate from gray matter atrophy. Future research may either identify separate mechanisms for early GM and WM change, or else show a common underlying process, which at the early stages manifests itself more strongly in the fornix–fSROI association.

#### **ACKNOWLEDGMENTS**

Thanks to Anna Lam and Evan Hamaguchi for data collection and quality control and discussion of experimental approaches in this article.

#### **REFERENCES**

Acosta-Cabronero, J., Williams, G. B., Pengas, G., and Nestor, P. J. (2010). Absolute diffusivities define the landscape of white matter degeneration in Alzheimer's disease. *Brain* 133(Pt 2), 529–539. doi:10.1093/brain/awp257


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 30 January 2014; accepted: 15 May 2014; published online: 28 May 2014. Citation: Fletcher E, Carmichael O, Pasternak O, Maier-Hein KH and DeCarli C (2014) Early brain loss in circuits affected by Alzheimer's disease is predicted by fornix microstructure but may be independent of gray matter. Front. Aging Neurosci. 6:106. doi: 10.3389/fnagi.2014.00106*

*This article was submitted to the journal Frontiers in Aging Neuroscience.*

*Copyright © 2014 Fletcher, Carmichael, Pasternak, Maier-Hein and DeCarli. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Diffusion tensor imaging in Alzheimer's disease: insights into the limbic-diencephalic network and methodological considerations

# *Julio Acosta-Cabronero\* and Peter J. Nestor*

*Brain Plasticity and Neurodegeneration Group, German Center for Neurodegenerative Diseases (DZNE), Magdeburg, Germany*

#### *Edited by:*

*Kenichi Oishi, Johns Hopkins University, USA*

#### *Reviewed by:*

*Marco Bozzali, Fondazione Santa Lucia, Italy Paul Gerson Unschuld, University of Zürich, Switzerland*

#### *\*Correspondence:*

*Julio Acosta-Cabronero, Deutsches Zentrum für Neurodegenerative Erkrankungen (DZNE), Universitätsklinikum Magdeburg, Leipziger Strasse 44, Haus 64, 39120 Magdeburg, Deutschland e-mail: julio.acosta@dzne.de*

Glucose hypometabolism and gray matter atrophy are well known consequences of Alzheimer's disease (AD). Studies using these measures have shown that the earliest clinical stages, in which memory impairment is a relatively isolated feature, are associated with degeneration in an apparently remote group of areas—mesial temporal lobe (MTL), diencephalic structures such as anterior thalamus and mammillary bodies, and posterior cingulate. These sites are thought to be strongly anatomically inter-connected via a limbic-diencephalic network. Diffusion tensor imaging or DTI—an imaging technique capable of probing white matter tissue microstructure—has recently confirmed degeneration of the white matter connections of the limbic-diencephalic network in AD by way of an unbiased analysis strategy known as tract-based spatial statistics (TBSS). The present review contextualizes the relevance of these findings, in which the fornix is likely to play a fundamental role in linking MTL and diencephalon. An interesting by-product of this work has been in showing that alterations in diffusion behavior are complex in AD—while early studies tended to focus on fractional anisotropy, recent work has highlighted that this measure is not the most sensitive to early changes. Finally, this review will discuss in detail several technical aspects of DTI both in terms of image acquisition and TBSS analysis as both of these factors have important implications to ensure reliable observations are made that inform understanding of neurodegenerative diseases.

**Keywords: neurodegenerative diseases, Alzheimer's disease neurobiology, axonal loss,** *circuit of Papez***, long association tracts, splenium, DTI criteria, Alzheimer's disease biomarkers**

## **INTRODUCTION**

Alzheimer's disease (AD) is characterized, histopathologically, by amyloid deposition and neurofibrillary tangles (composed of hyperphosphorylated tau); these features occur in somewhat topographically distinct distributions in the brain. The ultimate outcome of AD is neuronal loss though exactly how histopathological features interact with each other and how they relate, in turn, to neuronal degeneration remains rather unclear. It seems reasonable to assume, nonetheless, that neuronal—and synaptic—loss is critical to the development of cognitive impairment. To date, therapeutic attempts to slow the course of AD have targeted the histopathology. A possible alternative, or even complimentary approach, might be to understand what makes some neuronal populations in the central nervous system more vulnerable to degeneration than others with a view to finding ways to make neurons less vulnerable to pathological insult—regardless of what that pathological insult might be. Such an approach requires a precise understanding of the spread of neurodegeneration and this can only be achieved by studies in humans rather than disease models.

Historically, *in vivo* work to understand the landscape of neurodegeneration in AD has focused on atrophy detection using T1-weighted magnetic resonance imaging (MRI) and metabolic studies using (18F)-2-fluoro-deoxy-D-glucose positron emission tomography (FDG-PET)—the latter being primarily a marker for synaptic loss. Both of these modalities typically focus on changes in gray matter (GM) and have tended to show what, at first glance, appear to be different profiles of degeneration (Ishii et al., 2005). Structural MRI has been good at highlighting mesial temporal lobe (MTL) atrophy (Du et al., 2001; Jack et al., 2002, 2004), while FDG-PET has typically highlighted early changes in posterior association cortex (Minoshima et al., 1995) with the posterior cingulate region in particular appearing as the first hypometabolic region in very early symptomatic AD (Minoshima et al., 1997; Nestor et al., 2003a). This apparent discrepancy between structural MRI and FDG-PET may well, however, be technical. For instance, much of the work on MTL atrophy derives from region-of-interest (ROI) studies that did not examine for atrophy elsewhere. Where whole brain analyses were conducted, most used the voxel-based morphometry (VBM) technique (Ashburner and Friston, 2000) though recent evidence highlights that although this method is good at identifying MTL atrophy, it is relatively insensitive to atrophy in the isocortical ribbon (Diaz-De-Grenu et al., 2014); in contrast, using the Freesurfer's cortical thickness method (Dale et al., 1999; Fischl et al., 1999), atrophy of the cortical ribbon in posterior association cortex emerges in a pattern highly reminiscent of that seen with FDG-PET (Diaz-De-Grenu et al., 2014). Similarly, although ROI studies in the cortical ribbon are rare, manual volumetry of the posterior cingulate region in MCI-stage AD has confirmed comparable degrees of atrophy to that seen in the hippocampus (Choo et al., 2010; Pengas et al., 2010). Looking at the apparent discrepancy in these imaging modalities from the FDG-PET side, voxel-based analysis (VBA) appears relatively insensitive to MTL changes in early AD; however, using MRIderived ROIs to calculate cerebral metabolic rates (i.e., quantified imaging), FDG-PET identified significant hypometabolism not only in the posterior cingulate but also the MTL, as well as anterior thalamus and mammillary bodies (Nestor et al., 2003b). This "limbic-diencephalic" network, therefore, appears to be the correlate of very early symptomatic AD when memory impairment is a relatively isolated feature. It is interesting, therefore, that these network structures—MTL, posterior cingulate, anterior thalamus and mammillary bodies—which are connected through the *circuit of Papez* (Papez, 1937), have all been individually implicated from focal lesions in human amnesia.

The hypothesis that these structures are degenerating in concert predicts that white matter (WM) projections between these areas—such as, for instance the fornix, which links the MTL with the diencephalon—should also show signs of degeneration. To test such hypotheses as well as to understand how degeneration in AD may impact on areas remote from GM degeneration more generally, the relatively more recent technique of diffusion tensor MRI offers considerable promise. Diffusion MRI enables mapping of WM microstructure alterations in development, aging and neurological disorders, and has therefore become an important tool in the study neurodegeneration. Parenchymal WM is composed of bundles of axons (or fiber tracts) that interconnect GM areas. The diameter of neuronal fibers is well below MRI resolution but the technique can be sensitized to measure the displacement of water molecules as a surrogate marker of tract integrity. Axonal membranes, myelin sheaths and cytoskeletal constituents such as microtubules and neurofilaments are long structures that may hinder water diffusion preferentially perpendicular to their length; this phenomenon enables MRI to detect abnormalities caused by neurodegenerative diseases such as loss of fibers, demyelination, damage within fibers or to support tissue around them (Beaulieu, 2002). The technique remains fairly new and there has been a relatively steep learning curve in terms of understanding and interpreting diffusion tensor imaging (DTI) (Alger, 2012; Winston, 2012); this learning process is far from over. What is clear at this time is that technical factors—both in terms of acquisition parameters and analysis methods—play a more important role in terms of generating spurious findings compared to older modalities such as structural T1-weighted imaging and FDG-PET. The outcome being that the DTI literature in AD can come across as a confusing jumble of inconsistent abnormalities—an unsurprising observation considering DTI's vulnerability to spurious results when suboptimal experimental designs are employed. The prescription of DTI acquisitions with fewer-than-recommended diffusion-encoding directions, low *b*-values or thick slices; the study of mild cognitive impairment (MCI) cohorts without clinical outcome and the detrimental effect of Gaussian smoothing in post-processing pipelines are the most likely contributors to such inconsistencies.

It is crucial therefore that clinicians and researchers have some understanding of the pitfalls that can arise from inadequate methods in order to interpret what have often been rather contradictory results in DTI studies. This review will focus on what we have learned from DTI in AD to date, but also will go into some detail in explaining the methodological issues that can cause problems for DTI studies.

# **THE PHYSICAL BASIS OF DTI**

The present section is not intended as a comprehensive description of DTI theory—for that, extensive reviews can be found elsewhere (Kingsley, 2006a,b,c). It aims, however, to provide an intuitive understanding of the method and an appreciation of the possible tensor dynamics that can arise with neurodegeneration; these are important theoretical considerations because diffusion, as measured with the "single tensor" model (Basser et al., 1994b), is based on a number of assumptions that lead to certain limitations; limitations that have been extensively discussed in the diffusion MRI literature (Jones and Cercignani, 2010; Jones et al., 2013b), but that are often overlooked in clinical studies. Nevertheless, DTI is a powerful technique to study neurodegenerative diseases; thus, an overview of its theoretical foundations will also enable a more clear understanding of the discoveries that have been made to date in AD. The theoretical underpinning of DTI can, however, seem impenetrable for nonspecialists even when most of the background mathematics is omitted, as is the case in this section. For readers, therefore, who might find the physics described in this section too daunting, we suggest simply focusing on the Figures and their captions; and with a brief understanding of what a diffusion tensor means, one can then skip on to the following section "Technical considerations for clinical DTI studies" for a discussion of the factors that influence the suitability of DTI scans for clinical studies.

Brownian motion, also known as "self-diffusion," is a physical phenomenon arising from matter's intrinsic thermal energy that leads to pseudo-random kinetic fluctuations. Such molecular thermal motion is named after Robert Brown—a Scottish botanist—who first observed such behavior in grains of pollen suspended in water (Brown, 1828). A full mathematical description emerged a few decades later when Adolf Fick—a physiologist studying mass transport in saline solutions—proposed that microscopic motion could be seen as a probability density function (*pdf*) of a particle's location in space and time, i.e., as the likelihood of finding a particle in a certain place at a certain time, and then went on to predict that self-diffusion must be governed by such time-dependent probabilistic behavior (Fick, 1855). Soon after the turn of the 20th century, Albert Einstein solved the general case of Fickian diffusion to provide an analytical answer to the practical question: *how far do "free particles" travel "in average" during a time interval?* Einstein proposed a simple "random walk" model comprising a series of discrete, unrestricted, independent and uncorrelated steps, which led him to discover that for such boundary conditions, Fick's *pdf* is a Gaussian distribution whose width—a measure of average molecular displacement—is modulated by two factors: time and a parameter dependent on fluid viscosity and temperature; he figured the latter must be a self-diffusion coefficient, *D* (Einstein, 1905) (see **Figure 1A**).

Since the inception of "spin echoes" in 1950, the MR signal has been known to be sensitive to molecular thermal motion (Hahn, 1950; Carr and Purcell, 1954; Torrey, 1956), which hereafter will be assumed to be that of hydrogen protons in water molecules within human brain tissue. Applying Einstein's relationship, it was determined that "free diffusion" under the effect of a steady magnetic field "gradient" attenuates the MR signal. This was first observed because field gradients (i.e., gradual increments or decrements in magnetic field strength) occur naturally when an experimental sample is introduced into the uniform magnetic field of a magnet and disturbs it. Magnetic field gradients, however, can be applied artificially to vary the main field gradually along one direction; such gradual field change, in essence, labels protons according to their spatial position. Stronger gradients, therefore, have the capability of giving a wider range of spatial signatures that result in higher image resolutions (if we refer to imaging gradients) or finer sensitivity to motion (if we refer to diffusion MRI). It was soon realized, however, that measuring self-diffusion through the application of sustained gradients was inconvenient and largely ineffective. Then in 1965, Stejskal and Tanner proposed that the amount of diffusion weighting in the MR signal could be finely controlled with a pair of identical, short-lived, magnetic field gradients (Stejskal and Tanner, 1965) (see **Figure 1B**). Pulsed gradients sequentially label, and then unlabel, protons according to their position; with the result that if a proton moves to a different location after a "diffusion time" i.e., when the second gradient is applied—, the proton gets assigned a wrong label and returns less signal according to how far it has moved. This principle—the basis for most diffusion MRI acquisitions today—results in an overall signal intensity attenuation due to free diffusion that depends on the self-diffusion coefficient, *D*, and the gradient characteristics. For simplicity, and

**FIGURE 1 | Measurement of the "self-diffusion" coefficient with magnetic resonance. (A)** Self-diffusion is a special case of diffusion that describes the relentless motion of microscopic particles in a substance. Such motion is random—it can be described by a "Gaussian" distribution whose width determines how far particles travel on average after some time; the important thing to remember is that only time, the "diffusion time," and the "self-diffusion coefficient," *D*, modulate this width. **(B)** Illustration of a magnetic resonance pulse sequence that is sensitive to self-diffusion: (i) an "excitation pulse" gives some energy to water protons; (ii) then the first "gradient" *encodes* where they are; (iii) and some (diffusion) time later, a

second gradient *decodes* their position by giving them a "signal penalty" according to how far they have moved away from their original position; (iv) as a result of this penalty, protons that have moved furthest, return the least signal. The overall signal attenuation is a function of *D* and the gradient characteristics—i.e., intensity (*G*), duration (*tG*) and diffusion time (*tD*)—usually conglomerated into a single term known as the "*b*-value." **(C)** Visual representation of the linear fit required to quantify *D* using two signals: a diffusion weighted measurement (i.e., with a non-zero *b*-value), *S*, that must be normalized by a measurement without diffusion weighting, *S*<sup>0</sup> (also often called b0); *D* can be inferred as the "negative" of the slope.

because in MRI there is a complex interaction between diffusion and imaging gradients, the overall effect of motion-sensitizing gradients is often synthesized into a so-called "*b*-value." This is convenient because the signal attenuation due to diffusion can be expressed simply by an exponential function of *D* and *b*, i.e., *S* ∝ exp (−*bD*). As in any spin echo experiment, however, the signal also decays due to transverse relaxation (T2) effects; thus, for the signal to depend only on the effects of water mobility, it must be normalized by a reference measurement without diffusion weighting, *S*0, also known in DTI jargon as a *"*b0 scan." Two measurements, therefore, are sufficient to infer a diffusion coefficient as: *D* = −ln (*S*/*S*0) /*b*; or in visual representation terms, that *D* is the negative of the slope connecting the two data points in **Figure 1C**.

The information that a Stejskal-Tanner experiment provides, however, is limited because the positions of resonant protons can only be encoded along the direction of the applied field gradient; thus, *D* only reflects the effect of self-diffusion along that specific spatial dimension. This is enough to characterize an isotropic medium, e.g., if measuring water's self-diffusion in a bucket, because diffusion measured in a given dimension will be same as that of every other dimension, but to probe a complex system with multiple, rotationally variant diffusion behaviors such as biological tissue one needs, in theory, an infinite number of observations with sensitizing gradients along an infinite number of diffusion orientations.

In an attempt to capture such directional dependency while keeping experimental requirements feasible, Basser et al. proposed the generalization of Einstein's equation by extending the Gaussian *pdf* idea to a second order, symmetric, definitepositive covariance matrix, **D**, and this is the diffusion tensor (Basser et al., 1994a):

$$\mathbf{D} = \begin{pmatrix} D\_{\text{xx}} \ D\_{\text{xy}} \ D\_{\text{xz}} \\ D\_{\text{yx}} \ D\_{\text{yy}} \ D\_{\text{yz}} \\ D\_{\text{zx}} \ D\_{\text{zy}} \ D\_{\text{zz}} \end{pmatrix} \tag{1}$$

**D** is a reciprocal matrix with *Dxy* = *Dyx*, *etc*—i.e., it only has six "independent" scalar elements: *Dxx*, *Dxy*, *Dxz*, *Dyy*, *Dyz*, and *Dzz*; thus, in order to reconstruct a diffusion tensor, the minimum requirements are one *<sup>S</sup>*<sup>0</sup> (*<sup>b</sup>* <sup>=</sup> 0 s/mm2 or b0) measurement and six diffusion measurements applying gradients along six noncollinear orientations (see **Figure 2A**)—i.e., Nd, the number of diffusion gradient directions, must be at least 6, though Nd can be larger with obvious benefits (see **Figure 2B**). The single tensor model proposed by Basser et al. requires that all diffusion directions (b-vectors), *b*-values and *b*<sup>0</sup> information (which can also be provided through multiple scans to improve stability) must be stored as a **B**-matrix, leading to a linear system of equations that can be solved for the six independent terms of **D** and for *S*<sup>0</sup> efficiently across an entire imaging volume.

Another convenient feature of the diffusion tensor is that it can be expressed in its quadratic form as an ellipsoidal surface that characterizes the water displacement probability at a given diffusion time, i.e., an ellipsoid that visualizes the 3D character of water mobility; though in the present form, i.e., that in

**FIGURE 2 | The "diffusion tensor." (A)** Six combinations of gradients applied with different intensities along three orthogonal orientations (*x*, *y*, and *z*), and a measurement without diffusion weighting, are the minimum requirements to probe diffusion in a 3D space; each non-collinear combination is called a "diffusion direction." **(B)** A larger number of diffusion directions, Nd, can estimate the directional dependence of microstructural restrictions to water diffusion with greater precision. **(C)** Basser et al. (1994b) proposed the formulation of a diffusion tensor (represented by an ellipsoid) to capture diffusion behaviors in 3D with a

finite Nd. **(D)** The tensor can be rotated to match the principal direction of diffusion, i.e., in white matter, the first eigenvector (ε1) matches the predominant fiber tract orientation; thus, the largest eigenvalue (λ1, known as "axial diffusivity") quantifies water mobility along this orientation. Analogously, ε<sup>2</sup> and ε<sup>3</sup> define the plane perpendicular to the axonal arrangement so that λ<sup>2</sup> and λ<sup>3</sup> can be averaged as a "radial diffusivity" (RD). In addition, all three eigenvalues can be averaged to estimate the "mean square displacement" of water molecules—that is, the "mean diffusivity" (MD).

Equation (1), **D** is still rotationally variant because it depends on the principal diffusion orientation relative to the *x*, *y*, and *z* axes of the scanner or to whatever coordinate system the b-vectors were in (see **Figure 2C**). **D**, however, is positive and symmetric regardless of such orientation; thus, using linear algebra, it can be easily made diagonal for a specific orientation—i.e., it can be decomposed into a set of orthonormal eigenvectors and related eigenvalues (see **Figure 2D**):

$$\mathbf{D} = \begin{pmatrix} \varepsilon\_1 \\ \varepsilon\_2 \\ \varepsilon\_3 \end{pmatrix} \begin{pmatrix} \lambda\_1 & 0 & 0 \\ 0 & \lambda\_2 & 0 \\ 0 & 0 & \lambda\_3 \end{pmatrix} \begin{pmatrix} \varepsilon\_1 \ \varepsilon\_2 \ \varepsilon\_3 \end{pmatrix} . \tag{2}$$

In essence, the diagonalization step in Equation (2) rotates the principal axes of **D** to match the principal directions of diffusivity, leading, as a result, to a diffusion tensor that is rotationally invariant for each imaging "voxel."

A number of metrics can be derived from the diffusion tensor; for example, the 3D "mean square displacement" of water molecules during a diffusion time can be derived by averaging the three tensor eigenvalues to form a DTI metric known as "mean diffusivity": MD = (λ<sup>1</sup> + λ<sup>2</sup> + λ3)/3. In addition, the rotational invariance of the diffusion tensor has clear advantages because water molecules can be generally assumed to "diffuse" due to their thermal energy more readily along the length of a uniaxial environment than perpendicular to it. It is, thus, commonly assumed when referring to, e.g., parenchymal WM, that the eigenvector, ε1, associated with the largest eigenvalue, λ1 known as "axial diffusivity"—maps the principal orientation of a fiber tract. Similarly, diffusivities along ε<sup>2</sup> and ε3—typically averaged to form "radial diffusivity," i.e., RD = (λ<sup>2</sup> + λ3)/2—reflects the diffusion behavior transverse to an axonal path. Therefore, the inter-relationship between eigenvalues—i.e., between axial and radial diffusivities—contains relevant information about the geometry of the restricting microstructure. For example, in a scenario where axons are tightly packed together such as, e.g., the mid-sagittal corpus callosum, water molecules are more restricted perpendicular to the axons, hence λ<sup>1</sup> >> λ<sup>2</sup> ≥ λ3, or λ<sup>1</sup> >> RD—i.e., it can be represented as a cigar-shaped diffusion ellipsoid with the long axis parallel to the axons (see **Figure 3**). In other regions, however, where, e.g., two tightly packed WM bundles cross, water molecules in the extra-cellular space might travel more freely across a plane, λ<sup>1</sup> ≥ λ<sup>2</sup> >> λ3, hence λ<sup>1</sup> > RD i.e., the ellipsoid adopts a planar geometry; whereas in WM areas where multiple fiber bundles cross, or in GM or cerebrospinal fluid (CSF), where water diffusion does not exhibit a preferential orientation because obstacles are randomly distributed in space or because there are no restrictions at all, the ellipsoid is largely "isotropic," i.e., λ<sup>1</sup> ≈ λ<sup>2</sup> ≈ λ3, thus λ<sup>1</sup> ≈ RD. The eccentricity of the diffusion ellipsoid, therefore, is an important property that can provide useful information about the biological tissue under investigation. This property is typically measured using the "second moment" of the diffusion tensor, because of its robust noise properties, by a metric known as fractional anisotropy (FA): FA <sup>=</sup> <sup>√</sup>3/<sup>2</sup> (λ<sup>1</sup> − MD) <sup>2</sup> <sup>+</sup> (λ<sup>2</sup> <sup>−</sup> MD) <sup>2</sup> <sup>+</sup> (λ<sup>3</sup> <sup>−</sup> MD) 2 / λ1 <sup>2</sup> <sup>+</sup> <sup>λ</sup><sup>2</sup> <sup>2</sup> <sup>+</sup> <sup>λ</sup><sup>3</sup> 2 (Basser and Pierpaoli, 1996). FA can be seen as the relative ratio

membranes are thought to be the main restricting boundary to water mobility in white matter. As such, the ratio of axial to radial diffusivities, commonly described by a metric known as "fractional anisotropy" (FA), can reflect the coherence (sometimes thus the integrity) of packed axons in white matter. FA approaches one in well-organized tracts, where the diffusion ellipsoid is elongated along the principal tract orientation; and tends to zero, in less coherent environments—i.e., in heterogeneous areas of many crossing fibers, and in gray matter or cerebrospinal fluid—, where the ellipsoid resembles a sphere. In the full complexity of white matter microstructure, however, the ellipsoid can take a wider range of geometries from that of a rugby ball to that of surfboard or a dish—with the degree of sphericity and planarity dictated primarily—but not only—by axon packing density, degree of myelination and/or the geometrical arrangement of crossing, kissing and/or splaying fiber populations.

between axial and radial diffusivities, where its boundary situations are: (i) *FA* = 0 for a perfectly "ergodic" or "isotropic" behavior, i.e., when λ<sup>1</sup> = RD, and (ii) *FA* = 1 for 1D molecular displacements in a 3D space, i.e., when RD is infinitesimally small relative to λ1. The latter behavior is the extreme case for "anisotropic" diffusion. In graphical terms (**Figure 3**), isotropic diffusion (*FA* = 0) is a perfect sphere because it means water can travel equally in all directions, whereas if water mobility is restricted disproportionally along one or two dimensions in space, FA moves away from zero.

The pseudo-random (Gaussian) nature of unrestricted diffusional motion is, in turn, the theoretical basis supporting the single tensor model in DTI (Basser et al., 1994b); this highlights, however, one of DTI's major limitations—it is based on an idealization. As already discussed, in WM, cytoarchitectural barriers hinder and physically restrict water mobility (see **Figure 4A**); such nuisances to self-diffusion, therefore, impose a different set of boundary conditions to Fick's equation, making "walks" no longer random. Instead, they become correlated and dependent on the geometry of the restricting boundaries; thus leading to *pdf* departures (narrowing) from the ideal Gaussian behavior (see **Figure 4A**). In such scenarios, the diffusion tensor proposed by Basser et al. does not contain enough terms to account for the higher order effects in non-Gaussian molecular displacement distributions; thus, they lead to the violation of DTI's basic assumption that the signal decay due to diffusion follows a single exponential behavior. Consequently, the single tensor model must be regarded as an approximation (**Figure 4B**); and the three tensor-eigenvalues—if they describe restricted water motion are not true self-diffusion coefficients, but "apparent" measures of diffusion or "diffusivities."

A further caveat is that most DTI experimental designs assume that tensor behaviors are independent of diffusion time—i.e., that the brain parenchyma is a restricted environment in a "pseudo-Gaussian" state, where diffusion time is long relative to the time needed for water molecules to be hindered/restricted by cellular membranes or other microstructural components. Note that the former are thought to be the primary microstructural restrictions to water diffusion in white matter (Beaulieu and Allen, 1994). This can be generally assumed to be valid because in typical clinical situations, diffusion times range between 30 and 50 ms, which translate to *in vivo* 1D molecular displacements of 13–17μm. Such displacements are an order of magnitude greater than, for example, the typical axon diameter in the corpus callosum of young individuals, which is approximately 1μm (Aboitiz et al., 1992)—though axons can also be much larger. In lay terms, the reason this is important is that in order to know that restrictions of water diffusion due to barriers in neural tissue are present, one must measure long enough "diffusion paths" to ensure water molecules have had a chance to "bounce off" some barriers. **Figure 5A** illustrates this scenario, where such paths are long relative to the distance between physical obstacles, enabling DTI to sense differential bulk behaviors parallel and perpendicular to white matter tracts. In patients with neurodegenerative diseases, however, cellular membranes may become more permeable; or the extra-cellular space might be greater due to axonal loss, demyelination, and/or glial pathology. Such effects—alone or in combination—may result in local environments where, during a given diffusion time, water molecules interact only partially with microstructural boundaries. **Figure 5C** illustrates such a scenario where some paths (e.g., the blue path) do not encounter any obstacle during a given diffusion time, i.e., they behave as if they were in an unrestricted medium. In theory, during disease progression the loss of large myelinated axons, or clumps of adjacent tracts, may result in diffusion paths that cannot capture the full extent of such "voidage." This is the so-called "quasirestricted" regime, in which measurements depend on diffusion time—i.e., if water molecules were allowed to diffuse twice as far by doubling the diffusion time, microstructural boundaries would be sensed again, resulting in smaller radial diffusivity relative to axial diffusivity, hence leading to a more anisotropic tensor. Such behavior may inadvertently compromise whole brain

**FIGURE 4 | The approximate nature of the "single tensor" model. (A)** "Restricted" diffusion processes such as those in white matter result in displacement probabilities that are no longer accurately described by a Gaussian "bell." **(B)** Therefore, the assumption that the signal attenuation must have a linear relationship in a logarithmic scale with the *b*-value is no longer valid, resulting in tensor diffusivities that are "apparent" (i.e., model approximations) rather than "true" measures of restricted self-diffusion.

assessments because the same neurodegenerative process, e.g., axonal loss, could impact diffusivities differentially across WM tract environments with different axon size distributions, packing arrangements or degrees of myelination. Suppose, for instance, an enlargement of the extra-cellular space of 20% due to axonal loss; its impact on RD would be different, for example, in a scenario where axons are lost sparsely than if a small bundle of large myelinated axons degenerate together. In the latter case, if molecular displacements were small relative to the "new" boundaries, FA would be underestimated as a result. It may be desirable to detect such diffusion-time dependencies (Baron and Beaulieu, 2014); but they imply that whole brain analyses would be biased toward such phenomena if they were proven to be present in some

**FIGURE 5 | The "diffusion regimes." (A)** White matter is a complex—but relatively ordered— microstructure primarily composed of nerve fibers (axons) and glial cells. Axons are bundled together; their main role is to transport substances intra-cellularly through microtubules and conduct electricity to enable inter-communication between cells. The myelin sheath—a "fatty" insulating layer around the axons—facilitates such conduction. In the healthy brain, such microstructure exerts restrictive boundary conditions to water diffusion. **(B)** The exact mechanism by which microstructural damage occurs in neurodegenerative diseases is unknown but it is conceivable that, after a period of instability, demyelinative and other axon degeneration processes will lead to longer "diffusion paths." If therefore, the diffusion time is sufficiently long and the gradients

sufficiently strong, such diffusion behavior will yield a change in signal attenuation that will be reflected in tensor diffusivities. **(C)** If, however, some diffusing molecules cease to interact with microstructural barriers during a given diffusion time, the axial to radial relationship would be dependent on the local geometry leading to heterogeneous tensor behaviors across the brain. **(D)** In the extreme case, where the diffusion time is too short for molecular displacements to be hindered, tensor diffusivity measures would be unable to detect further change. Large *b*-values—enabling long diffusion times and strong gradients—make the diffusion measurement with magnetic resonance both more sensitive to subtle microstructural alterations in highly restricted environments, and less prone to diffusion-time dependencies.

diseased regions but not in others. Furthermore, there might be scenarios of extreme degeneration, or CSF contamination in a voxel, where the majority of water molecules behave as in an unrestricted medium; under such circumstances, further loss of microstructural integrity would lie largely undetected with DTI because water mobility would not reflect the presence of any boundary (as in **Figure 5D**).

While these important theoretical considerations and limitations should not be ignored, DTI can be very useful to probe abnormal tissue microstructure. For instance, in a pseudo-Gaussian scenario where several fiber tracts are damaged as a consequence of a disease, diffusion hindrance would be reduced (see **Figure 5B**), i.e., *pdf*s would approximate more readily to a Gaussian distribution, and would return greater diffusivity values closer to the intrinsic unrestricted self-diffusion coefficient; it is this change in diffusivity pattern that can indicate abnormality in degenerative disease states. In addition, because the tensor model yields three rotationally invariant—and orthogonal—diffusivities (λ1, λ2, and λ3), a disproportionate diffusivity change along a particular orientation will impact on absolute (axial and radial) diffusivities, or composite measures of diffusion anisotropy such as FA as illustrated in **Figure 3**.

It is presently unknown whether diffusion-time dependencies might be a confounding force in the study of neurodegenerative diseases; for now, they remind us that caution must be at the forefront in DTI interpretations. Unfortunately, DTI studies in neurodegenerative diseases have often been guilty of overinterpreting results, particularly when such interpretations are based on consistency with prior knowledge rather than offering a proof of a given mechanism. Mechanisms underpinning DTI changes in clinical cohorts are often inferred by citing homology with controlled animal experiments or computer simulations, where processes such as myelin loss, fiber reorganization, changes in membrane permeability, *etc*, are modeled in isolation—for a extended review see Beaulieu (2002). Reductionist approaches such as animal experiments are important in that they can highlight the types of neural changes that are possible to detect with DTI but there is a potentially serious error in logic when such observations are exported directly to explain changes observed in human neurodegeneration. For instance, if a given DTI change has been observed in an animal model of demyelination and then the same DTI change is observed in a degenerative disease, one will often read that the mechanism in the latter is also demyelination. The error in logic is that there are only a finite number of ways in which the diffusion tensor can change, and for any given change, it is incorrect to assume that only one mechanism might be the explanation—there may be many different types of pathological change that could cause the same diffusion tensor alteration. Using a clinical analogy, this is the same error in logic as stating that because sub-arachnoid hemorrhage can cause headache, any person with a headache must have a subarachnoid hemorrhage. Furthermore, it seems highly unlikely that a single mechanism will dominate the diffusion weighted MR signal because fiber degradation processes such as Wallerian degeneration—which typically involve the complex combination of cytoskeleton dissolution, contraction, fragmentation, disintegration of myelin and glial disposal (Coleman and Freeman, 2010)—are unlikely to occur simultaneously across all cells in a single measurement voxel. Such complexity, worsened by the diffusion regime uncertainty, suggests that DTI results should not be interpreted beyond their canonical form—i.e., that increased (or reduced) diffusivities are caused by less (or more) restrictive barriers to water diffusion, and that diffusion anisotropy alterations are driven by orientation dependent changes in the configuration of such microstructural arrangement. What these exactly represent in neurobiological terms can only be speculated. Nevertheless, DTI-derived diffusivity and anisotropy measurements can offer important mechanistic insights toward understanding the sequence of events that lead to neurodegeneration by highlighting tracts that are diverging from normality, even if diffusion MRI, at present, cannot precisely state what the underlying pathology driving such divergence might be.

# **TECHNICAL CONSIDERATIONS FOR CLINICAL DTI STUDIES**

Historically, T1-weighting has been the MRI contrast of choice to study structural abnormalities in the human brain. It is clear that imaging parameter discrepancies or different field strengths result in differential method sensitivity, but overall they tend to be relatively concordant if the same processing steps are used. This is because structural MRI post-processing methods typically concentrate on resolving and standardizing different types of tissue using large amounts of prior anatomical knowledge; this makes such analysis strategies relatively immune to differences, e.g., in signal-to-noise ratio (SNR). Unlike structural MRI, however, DTI relies on *quantitative* information to yield meaningful assessments; thus, DTI sensitivity and stability strongly depend on how robustly the signals have been acquired. Therefore, additional precaution must be taken when reviewing the literature because the ability of diffusion MRI to accurately reconstruct tensorial information strongly depends on measurement SNR, which varies throughout the brain, between subjects and across studies. It is likely that when non-specialists read an imaging study, acquisition details—which may seem an impenetrable paragraph of technical jargon—are skipped over. Such information though is critical to DTI because some acquisition protocols that have been applied in clinical studies are simply not fit for purpose. It is beyond the scope of the present manuscript to discuss in detail every relevant factor—for a more specialized review, see e.g., Jones and Cercignani (2010)—but to help readers to critically evaluate the literature in this ever-growing field, a number of the most clinically relevant technical considerations will be briefly discussed:

#### **THE DTI ACQUISITION SCHEME**

The "*b*-value" is the parameter that tunes how much microstructural information the diffusion MR signal can carry. Relatively large *b*-values are needed to ensure both (i) that strong diffusion gradients sensitize water motion in complex, highly restricted environments, and (ii) that the diffusion time is sufficiently long to enable "meandering" diffusion paths to probe less restricted (or diseased) microstructure. Early computer simulations suggested that for a single non-zero *b*-value experiment, *<sup>b</sup>* <sup>=</sup> 1250 s/mm2 minimizes fitting errors for a mean diffusivity of <sup>0</sup>.<sup>7</sup> <sup>×</sup> <sup>10</sup>−<sup>3</sup> mm2/s, i.e., approximately that of brain parenchyma (Xing et al., 1997). More recent simulations have demonstrated that even taking the simplification that the human brain is constituted by single-orientation fiber populations, i.e., that crossing fibers are not present, the *b*-value—if interested in MD and FA should not be lower than *<sup>b</sup>* <sup>=</sup> 900 s/mm2 (Alexander and Barker, 2005). Such results highlight a pitfall with a large number of studies—particularly at 1.5 Tesla (T)—that used a maximum *b*value of 700 s/mm2. Capturing high order signal decays reflecting non-Gaussian restricted diffusion with a smaller-than-optimal single non-zero *b*-value translates to inaccurate "fits," which, as illustrated in **Figure 4B**, may lead to tensor reconstructions that do not accurately reflect the true diffusion process, and hence, the underlying microstructure. The use of larger *b*-values—closer to optimal for tensor estimation in WM—can help reduce such residual errors, but high *b*-values further attenuate the MR signal; thus, in order to ensure that diffusion-weighted signals are well above the "noise floor," a common practice is to take advantage of the higher baseline SNR at stronger magnetic fields (e.g., 3T MRI). It should be noted, however, that stronger fields are not always favorable for DTI because the diffusion-weighted signal is also T2 dependent, and T2 is shorter for stronger fields, which, in turn, translates to faster signal decays. This is undesirable, and poses a serious challenge, particularly at ultra-high field (7T and above), because whole brain DTI acquisitions typically consist of a single-shot echo planar imaging (EPI) pulse sequence with longlasting diffusion and readout gradients resulting in long delays before the signal can be acquired, i.e., long echo time. At 3T, however, the gain in baseline SNR usually overcomes the T2 penalty; that is why presently 3T MRI is the platform of choice for diffusion tensor neuroimaging. It is noteworthy, looking to the future, that low SNRs will become less of a problem for clinical studies with the advent of high performance gradient systems capable of stronger field gradients with shorter switching times, which enable the use of higher *b*-values with shorter echo times.

Returning to the present, it is important to avoid low SNR because the noise floor may damp signal decay, resulting in artefactually reduced diffusivities; this effect is orientation dependent in the WM because faster signal decays—caused by greater diffusivities along tracts—are more vulnerable to this effect, resulting in artefactual reductions of diffusion anisotropy. This is critically important because if one is attempting to map the distribution of WM change in degenerative brain disease it is possible to miss changes in affected areas where the local fiber orientations were more vulnerable to the effects of low SNR. Signal levels are modulated by the magnetic field strength, the type of radiofrequency coil used and imaging parameters such as the echo time (TE); number of *b*-values (N*b*); number of repeat excitations (NEX); receiver bandwidth or parallel imaging acceleration, e.g., GRAPPA (Griswold et al., 2002) or SENSE (Pruessmann et al., 1999). Unfortunately, however, SNRs are not usually reported in the literature. In addition, details about the number of channels available in phased-array coils, acceleration factor or receiver bandwidth are also often not reported in clinical studies, making comparisons between works hard to interpret from methods sections.

The influence of suboptimal gradient sampling schemes is also an important factor to consider when scrutinizing the literature. Diffusion weighting along six unique orientations (see **Figures 2A,B**) and the acquisition of a reference non-diffusion weighted (or b0) image are the minimum requirements to reconstruct DTI parametric maps; but to ensure primary diffusivities are independent of fiber orientation—i.e., to ensure DTI metrics are rotationally invariant –, one has to probe diffusion behaviors evenly and at a high spatial frequency. This is because tensor eigenvalues, hence diffusivities and anisotropy estimates, are more robust when WM tracts are collinear with one of the motion-sensitizing gradients (Jones, 2004). It is, therefore, essential that large numbers of diffusion-encoding orientations (Nd) are used for accurate DTI measurements in whole brain studies. In fact, the minimum number for robust tensor reconstruction has been estimated to range between *N*<sup>d</sup> = 20 (Papadakis et al., 2000) and *N*<sup>d</sup> = 30 directions (Batchelor et al., 2003; Jones, 2004), though a recent computer simulation has shown that a number of directions as low as *N*<sup>d</sup> = 12 can yield accurate MD estimates if the number of *b*-values is sufficiently large, e.g., *Nb* = 5 (Correia et al., 2009). Interestingly, such robust behavior with respect to noise could not be replicated using repeat measures (NEX = 5) with a single *b*-value. This work, therefore, highlights the importance of collecting more than one *b*-value—untrue for NEX > 1—both to sensitize the DTI measurement to a wider range of diffusivities—i.e., microstructural environments—and to reduce the measurement bias due to noise. In clinical research practice, however, the number of measurements, Nm, is typically limited to 60–70 for a total scan time of 10 min; hence an optimal compromise between Nd and Nb must be achieved. Correia et al.'s DTI simulation revealed that for *N*<sup>m</sup> = 60 measurements, if the interest is in MD, *N*<sup>d</sup> = 12 × *Nb* = 5 is the optimal set-up—i.e., a large number of *b*-values is preferred. For FA estimation, however, a larger number of directions—Nd ≥ 30 (as in Batchelor et al., 2003; Jones, 2004)—is needed to establish accurate axialto-radial diffusivity relationships, whereas little gain was observed for *N*<sup>b</sup> > 2. Therefore, an important outcome of these studies is that *N*<sup>d</sup> = 30 × *Nb* = 2 appears to be a good compromise to ensure that tensor reconstructions are reliable; leading to the notion that the majority of clinical studies—typically sensitized with a single *b*-value and often, *N*<sup>d</sup> ≤ 15 with NEX ≥ 2—may have led to poor measurement stability.

An additional factor that may play a role in DTI robustness is the accurate estimation of *S*<sup>0</sup> through multiple b0 scans (N*b*0). To our knowledge, this has not been systematically investigated, though in theory its impact will depend on SNR (TE, parallel acceleration factor, *etc*), and on the prescribed number of *b*-values—becoming less relevant with increasing N*b*. It is conceivable, however, that a large N*b*<sup>0</sup> can help stabilize the measurements for little additional scan time—particularly in single *b*-value experiments.

**Figure 6** illustrates the choice of *b*-value, N*<sup>b</sup>* and N*b*<sup>0</sup> dependence on DTI reconstructions. All maps were inferred from subsets (or the complete dataset) of a single acquisition with *<sup>N</sup>*<sup>d</sup> <sup>=</sup> 30, two non-zero *<sup>b</sup>*-values (*<sup>b</sup>* <sup>=</sup> 700 and 1000 s/mm2) and *Nb*<sup>0</sup> = 12—i.e., *N*<sup>m</sup> = 72. Notable improvement in the calculation of diffusivities and anisotropy can be observed with the increase in *b*-value from 700 to 1000 s/mm2. As would be expected, further improvements are noticeable when using the information from both *b*-values in combination; though the highest diffusivity-to-noise ratios are returned by the complete dataset with a larger number of b0 scans. It is, therefore, clear that they all contribute to improving DTI measurement stability.

Correia et al. (2009) showed that measurements from suboptimal schemes quickly deviate from true simulated values with decreasing SNR; but SNR is not only dictated by the factors enumerated above, but also by image resolution. One finds a plethora of voxel sizes in the published DTI literature. In the best case scenario, images are composed of isometric 2 × 2 × 2 mm3 voxels, but a large number of studies—particularly early research at 1.5T—used slices more than twice as thick as in-plane resolution to ensure sufficient measurement SNR. Non-isometric voxels, however, are undesirable for volumetric analyses because they encourage volume averaging along the slice direction, which translates to spurious direction-dependent error biases (Oouchi et al., 2007). An alternative would be to encode smaller matrices with larger isotropic voxels, which is common practice, but

**FIGURE 6 | Comparison of DTI parametric maps with different** *b***-value, number of** *b***-values (N***b***) and number of b0 scans (N***b***0).** MRI measurements were performed on a Siemens Verio 3T system (Siemens Medical Systems, Erlangen, Germany)—gradient coils capable of 45 mT/m and 200 T/m/s slew rate—with a 32-channel phased-array head-coil. Diffusion volumes were acquired using a standard twice-refocused, single-shot EPI pulse sequence: repetition/echo time = 9000/94 ms; matrix, 120 × 120; 63 contiguous slices aligned parallel to the anterior commissure/posterior commissure line; voxel size: 2 × 2× 2 mm3; 7/8-phase partial Fourier; bandwidth of 1667 Hz/pixel and echo

spacing of 0.68 ms. Diffusion gradients were applied along *N*<sup>d</sup> = 30 non-collinear directions (Siemens default vectors) with *Nb* = 2 non-zero *<sup>b</sup>*-values (*<sup>b</sup>* <sup>=</sup> 700 and 1000 s/mm2), and *Nb*<sup>0</sup> <sup>=</sup> 12 reference scans. Parallel imaging was enabled (GRAPPA, acceleration factor = 2 and 38 reference lines), leading to a total scan time of 11 min and 15 s. DTI maps were computed with standard tools from FSL's diffusion toolbox. (Left to right columns): (i) *Nb* <sup>=</sup> 1 (*<sup>b</sup>* <sup>=</sup> 700 s/mm2), *Nb*<sup>0</sup> <sup>=</sup> 1 (5:06); (ii) *Nb* <sup>=</sup> 1 (*<sup>b</sup>* <sup>=</sup> 1000 s/mm2), *Nb*<sup>0</sup> <sup>=</sup> 1 (5:06); (iii) *Nb* <sup>=</sup> 2 (*<sup>b</sup>* <sup>=</sup> 700 and 1000 s/mm2), *Nb*<sup>0</sup> <sup>=</sup> 1 (9:36); (iv) *Nb* <sup>=</sup> 2 (*<sup>b</sup>* <sup>=</sup> 700 and 1000 s/mm2), *Nb*<sup>0</sup> = 12 (11:15).

one should not ignore that partial volume effects caused by large measurement voxels can also be highly detrimental in DTI due to GM and/or CSF contamination. This, in turn, can lead to systematic errors in patient populations because typically patients have more brain atrophy than controls, and therefore, DTI alterations may reflect CSF inclusions rather than true microstructural changes. It should be noted that the WM tract of interest in the present Research Topic—the fornix—due to its close vicinity to ventricular CSF is particularly vulnerable to this phenomenon as illustrated in **Figure 7**.

In conclusion, although DTI is becoming a mature technique, there is not yet a universal agreement on the minimum requirements for a reliable DTI acquisition, or on an optimal image acquisition scheme for a given scan time, so we must reconcile with the fact that different DTI studies to date may have sensitized their acquisitions very differently.

#### **THE POST-PROCESSING METHODOLOGY**

Post-processing methods are also highly relevant to interpreting DTI literature. To date, a number of analysis strategies have been used including region-average, histogram or voxel-/cluster-based. No method, however, has demonstrated systematic superiority in every scenario; though it is widely accepted that for unbiased whole-brain assessments, the tract-based spatial statistics (TBSS) approach (Smith et al., 2006) has advantages and is the most desirable technique for volumetric DTI analysis at present (see **Figure 8**). TBSS—part of FMRIB's software library or FSL (Smith et al., 2004)—enables whole brain assessment of WM tract integrity in neurodegenerative diseases using DTI data without the need for *a priori* hypotheses about the spatial location of degenerative involvement. TBSS circumvents the lack of anatomical landmarks in WM, which is a limiting factor for manual tracing of tracts of interest in native space. TBSS deals with this problem by automatically co-registering all DTI parametric maps to a standard template; also, unlike histogram analyses, it can resolve the specific location of abnormal clusters; it is not biased by the dependency issues known to affect tractography based regional analysis (Besseling et al., 2012); and finally, TBSS does not require convolution with a smoothing kernel, as is used in standard VBA methods, due to the introduction of a skeletonisation step. The skeleton is derived from averaging FA images across all subjects and subsequent identification of tract centers;

**FIGURE 7 | White matter tracts that are usually prone to measurement error. (A)** The body of the corpus callosum (top arrow) and the body of the fornix (bottom arrow) are often problematic because they are adjacent to cerebrospinal fluid. The cingulum bundle, **(B)** which is typically less vulnerable at the level of the posterior cingulate, **(C)** is sometimes prone to

partial volume contamination and other types of measurement error in the parahippocampal region due to its thinner physical appearance (Jones et al., 2013a) and its closer proximity to rostral temporal areas known to be affected by magnetic susceptibility artifacts. Note in addition that orbitofrontal white matter is also vulnerable to such artifacts.

**FIGURE 8 | Tract-based spatial statistics (TBSS) processing pipeline.** (Left to right) DTI-derived fractional anisotropy (FA) images are co-registered to a template; they are then averaged, from which a "skeleton" containing all major tract centers common to all subjects is derived. Skeleton voxels with low *FA*-values (typically *FA* < 0.2) are excluded to ensure only white matter is present. Next, spatially normalized FA images are projected to the skeleton.

In this step, the center of each tract is identified for each individual FA image, and projection vectors to their analogous location in the skeleton are computed and applied. Transformation fields (to template space) and projection vectors (to the skeleton) are then also applied to the additional DTI parametric maps (i.e., MD, λ1, RD). Finally, non-parametric, permutation based, statistical testing of the null-hypothesis is performed.

this ensures the analysis is carried out exclusively on definite WM. But the main advantage of projecting DTI data to a WM skeleton is that it corrects, to some extent, co-registration inaccuracies, rendering the data more comparable than in VBA. In addition, performing statistics only along the center of major WM bundles minimizes the effects of partial volume contamination and reduces the number of statistical tests. Furthermore, TBSS' non-parametric statistical method does not require normally distributed data and, as for any randomization approach, it is inherently robust against Type-I errors (Winkler et al., 2014).

The default TBSS processing pipeline, however, suffers from its own sources of inaccuracy and limitations: it fails to rotate the **B**-matrix after realigning all images to account for motion and eddy-current effects (Leemans and Jones, 2009); it implements a fast—but oversimplified—tensor fitting routine (Jones and Cercignani, 2010) that assumes single fiber populations in each voxel, hence leading to potential miscalculations in areas of crossing and kissing fibers (Jbabdi et al., 2010); its non-linear coregistration algorithm, FNIRT, performs moderately relative to other *state-of-the-art* methods (Klein et al., 2009); assessments are spatially restricted to WM tract centers only—ignoring, therefore, a large amount of WM1 ; and finally, it suffers from spatial- and orientation-dependent statistical sensitivity (Edden and Jones, 2011). That is, TBSS is not completely unbiased. Such limitations should not be ignored; in fact, they suggest that in certain scenarios other methods might be preferable. For example, the fornix is often excluded from the WM skeleton due to the perennial issue of image misregistration (Keihaninejad et al., 2012) or does not survive correction for multiple comparisons at the cluster level due to its small physical size (Acosta-Cabronero et al., 2012; Schwarz et al., 2014). The fornix, however, has clear boundaries in FA maps (at least through the midline), and more advanced co-registration methods exist; thus, although FNIRT-based TBSS studies in AD will be the focus here to homogenize the potential confounding error introduced by different post-processing routines and because, to date, it is the most widely used method of whole-brain DTI analysis, it should be highlighted that manual tracing or automated regional extraction using high-performing, tensor-orientation enabled (Zhang et al., 2006) or intensitybased, diffeomorphic image registration algorithms (Avants et al., 2008) might be better suited to study structures such as the fornix.

Turning to the issue of image misregistration with standard FSL tools, a recent study has compared the warping performance for a number of co-registration strategies, including the standard TBSS procedure of warping each subject to the FMRIB's FA template (ST-TBSS), the previously recommended method of using the most representative subject as an intermediate registration step (RS-TBSS), and building study-specific (SS-TBSS) or group-wise templates (GW-TBSS)—all in the context of AD (Keihaninejad et al., 2012). The lack of ground truth makes the visual assessment of TBSS results in AD hard to interpret; however, using inter-subject variability and a proxy-measure of false positive rate as indicators of overall warping performance, Keihaninejad et al.'s results suggest that RS-TBSS is substantially inferior to the other methods, and that the use of the GW-TBSS approach might be beneficial.

#### **THE SUBJECT COHORT**

An important factor that must be considered is the nature of the patient cohort under investigation. Many studies focus on the prodromal stages of AD—typically patients diagnosed with subjective or mild cognitive impairment (SCI/MCI), genetically at risk (e.g., carriers of the apolipoprotein E ε4 allele), and/or those who are positive for an AD biomarker, e.g., CSF analysis or amyloid-ligand PET. It is noteworthy, however, that not all risk indicators are equivalent; while amyloid-PET is an unambiguous marker of AD neuropathology (Clark et al., 2012), CSF markers are not as robust due to inconsistent standardization (Verwey et al., 2009). In addition, genetic risk profiles (for sporadic AD) are problematic because they are only risk factors. Most troublesome, though, are studies that report cross-sectional cohorts of SCI or MCI with neither biomarker nor longitudinal outcome data to confirm symptoms were due to early AD. Cohorts without such information are impossible to interpret because of the potentially large number of false positive cases meeting such inclusion criteria (Jicha et al., 2006). This problem can be further compounded if imaging studies stratify MCI into amnestic (aMCI), non-amnestic (naMCI) and multi-domain (mdMCI), or, into "early" and "late" MCI without outcome or biomarker data. Individuals belonging to an mdMCI or late MCI group, being more cognitively impaired, have a higher probability of having AD, while naMCI has a lower probability (Mitchell et al., 2009). These are, however, only probabilities and do not indicate the pathology of each individual in a group. If the aim is to study the neurobiology—as it should be in an imaging study—of prodromal AD, a subject with naMCI due to AD is an appropriate inclusion whereas one with mdMCI that is not due to AD pathology is not. The present review, therefore, will not discuss such studies unless outcome data through clinical follow-up, or amyloid-PET/CSF analysis, confirmed that subjects in these pre-dementia stages were likely to have had incipient AD. It should be noted that meeting established criteria for probable AD (McKhann et al., 1984) does not completely predict AD neuropathology at autopsy, but the risk of including false positive cases is far lower than for SCI/MCI.

A further confounding effect that must always be considered is the comorbidity of WM hyperintensities. Studies typically deal with this by excluding cases with such lesions through visual inspection of T2-weighted images; this can also be supported by semi-quantitative measures using the Fazekas visual rating scale (Fazekas et al., 1987). It should be highlighted, however, that there is no consensus on what constitutes an unacceptable level of lesion burden for a DTI analysis in dementia although such information is very relevant given the extremely high prevalence of WM hyperintensities in the target population. This would be an interesting topic for further research, as would more automated

<sup>1</sup>In practical terms, however, this is not a problem for neurodegenerative diseases in which degeneration involves large-scale networks as opposed to tiny focal lesions. Visual inspection of a skeleton typically demonstrates that TBSS does not systematically miss any large chunks of the brain. In fact it can be desirable, particularly in neurodegenerative diseases, where the data reduction step to tract centers eliminates to a greater extent the vulnerability to partial volume contamination due to focal lesions.

and quantifiable measures of lesion load. For instance, it seems reasonable that subjects with negligible amounts of WM signal change would still be appropriate for inclusion in a DTI study, but this immediately raises the question of how one defines "negligible." More objective scales and a consensus on an acceptable threshold would, therefore, be desirable.

#### **MULTI-CENTER STUDY DESIGNS**

Multi-site designs are becoming increasingly common in AD research, particularly with the advent of the Alzheimer's Disease Neuroimaging Initiative (ADNI) and related spin-offs (Jack et al., 2010). Whilst such initiatives incorporate a pilot stage to estimate measurement stability across centers, a systematic assessment of inter-site DTI variability has not yet been carried out within the ADNI framework. An independent European initiative, however, performed such measurements and found significant FA variations across scanners, leading to the conclusion that FA variability cannot be fully controlled by the use of identical scanner type and acquisition parameters (Teipel et al., 2011). This result highlights that, at least until full cross-validations confirm otherwise, multicenter studies should account for potentially confounding scanner effects—a practice that is not commonly carried out. It should be noted, therefore, that multi-center studies in the present review have been included only to confirm widely reproduced observations, but little emphasis has been given to idiosyncratic results generated from such datasets.

#### **LITERATURE SYNTHESIS: INCLUSION CRITERIA**

In an effort to evaluate the DTI literature to date in AD, studies meeting the "essential" criteria for this review (**Table 1**) were selected. **Table 2** illustrates the iterative process that led to identification of 13 publications in which an acceptable acquisition and analysis protocol was employed; and in which patients had either clinically probable AD, MCI-stage AD where longitudinal follow-up or biomarkers were used to define probable AD status, or patients with autosomally inherited known gene mutations for AD (Acosta-Cabronero et al., 2010, 2012; Douaud et al., 2011; Bosch et al., 2012; Huang et al., 2012; Canu et al., 2013; Fieremans et al., 2013; Mahoney et al., 2013; Nir et al., 2013; Rowley et al., 2013; Ryan et al., 2013; Lim et al., 2014; Molinuevo et al., 2014). It should be stressed that **Table 2** highlights the order of iterations by which studies were excluded; it should not be read to mean that for each criterion the number of excluded studies equals that number of studies in total that failed on that specific criterion because many studies failed on multiple criteria. **Table 1** also includes a list of proposals of additional criteria for future studies based on current knowledge.

Details of the 13 studies that were identified from the literature review can be found in the Supplementary Table. It should also be noted that several of the studies (Douaud et al., 2011; Bosch et al., 2012; Huang et al., 2012; Nir et al., 2013; Rowley et al., 2013) included both AD and MCI groups where the latter did not

#### **Table 1 | Selection criteria for DTI studies in Alzheimer's disease included in the present review and additional guidelines for future studies.**

#### **Essential (inclusion criteria for this review)**

#### **DTI ACQUISITION**


#### **STUDY COHORT**



#### **DATA ANALYSIS**


#### **Desirable (additional suggested criteria for future studies)**

#### **DTI ACQUISITION**


#### **DATA ANALYSIS**

**+** Avoidance of RS-TBSS pipelines


#### **Table 2 | Summary of the hierarchical iterative steps leading to the selection of 13 DTI studies in Alzheimer's disease.**

meet entry criteria for the present review. Only data from the AD groups is included from such studies.

#### **TBSS IN AD: FINDINGS FROM THE LITERATURE OVERVIEW**

After careful literature filtering for potential technical problems in past studies, a number of consistent TBSS behaviors were identified; it would be wrong to assume, however, that such results provide a definitive understanding of WM changes in AD because, as discussed, alterations in structures that are prone to error may lie undetected for technical reasons. In contrast, prominent DTI effects that have been reproduced in a number of methodologically sound studies are evidence of real phenomena and that may, thus, be interpreted with scientific rigor as disease-related processes.

#### **WHAT DOES THE DISTRIBUTION OF DIFFUSION TENSOR ALTERATIONS TELL ABOUT AD?**

The selection process identified TBSS studies in AD that generally had a consistent common denominator—i.e., widespread and confluent tensor abnormalities in parietal, temporal and pre-frontal WM—specifically involving long association fibers (including limbic tracts) and inter-hemispheric connections through the corpus callosum. Generally speaking changes were more apparent in posterior areas compared to frontal areas as would be expected from prior knowledge of cortical atrophy studies and FDG-PET. **Figure 9** exemplifies such a distribution in a mild AD cohort (Acosta-Cabronero et al., 2012), where the posterior corpus callosum, cingulum bundle at the level of the posterior cingulate gyrus, superior longitudinal fasciculus and sagittal stratum (including the inferior longitudinal fasciculus, fronto-occipital fasciculus, and posterior thalamic radiation) were affected. Additional involvement is also commonly found in superior temporo-parietal WM at the level of the corona radiata and centrum semiovale—composed of intertwined long/short association and projection tracts.

Two exceptions, however, were found; the studies by Ryan et al. (2013) and Lim et al. (2014) did not find any cross-sectional differences in pre-symptomatic, autosomal dominant (PSEN1) mutation carriers and a group of MCI-stage AD patients defined by CSF biomarkers, respectively. The group reported by Ryan et al. (2013) were estimated to be a mean of 5.6 years from symptom onset so it simply may be that the group was too early to have major changes in WM; to this end it was notable that as a group there was also no significant hippocampal atrophy. The group only comprised *N* = 10 mutation carriers which also suggests it may have been underpowered to detect subtle changes with whole brain TBSS analysis; finally, the control group used was significantly older than the PSEN1 group which may have also masked abnormalities. The cohort in Lim et al. (2014) were *N* = 16 early MCI-stage AD subjects scanned under the umbrella of the ADNI2 framework, i.e., on 14 different scanners, which may, thus, have contributed to the lack of sensitivity. Though one might be tempted, in light of these results, to conclude that DTI is unable to detect WM changes in such early stages, the *post-hoc* regional study carried out by Ryan et al. (2013); the TBSS study by Molinuevo et al. (2014) in a group of cognitively normal and SCI subjects (positive for CSF biomarkers); and an atlas-based regional study (with suitable methods) in cognitively normal (amyloid-PET positive) subjects (Racine et al., 2014) suggest otherwise, as these reported DTI alterations that were consistent with the distribution in **Figure 9**.

The results from Canu et al. (2013) are also worthy of further discussion. The study examined two AD cohorts with different onset ages, but matched otherwise for disease severity, with two appropriately age-matched control groups. The early-onset group showed very widespread abnormalities highly concordant with **Figure 9**, but that spread slightly more anteriorly along association tracts and the genu of the corpus callosum. These were moderately impaired cohorts, i.e., average mini-mental cognitive examination (MMSE) score (Folstein et al., 1975) around 20/30; thus it would be reasonable to expect that changes had extended more extensively into frontal areas. Interestingly, however, the result in the late-onset cohort was relatively insensitive. Whilst differences were found in posterior areas largely overlapping the

**FIGURE 9 | TBSS in AD.** TBSS results for *N* = 43 early-stage AD patients (age: 70 ± 6, <MMSE>= 24 ± 4) vs. *N* = 26 matched controls (age: 68 ± 6) (Acosta-Cabronero et al., 2012). Pink clusters denote increased MD in patients, whereas those in yellow represent FA reductions at *P* < 0.05 enabling threshold free cluster enhancement (TFCE) (Smith and Nichols, 2009) and controlling the family-wise error (FWE) rate.

Thresholded statistical maps are overlaid onto the TBSS skeleton and MNI152 template. Sagittal MNI coordinates are given in millimeter where *x* < 0 is left. Abbreviations: [SLF] superior longitudinal fasciculus; [SS] sagittal stratum; [CR] corona radiata; [s/b/gCC] splenium/body/genu of the corpus callosum; [CGc/h] cingulum at the level of the posterior cingulate/parahippocampus; and [FX] fornix.

clusters in **Figure 9**, the distribution was markedly less extensive, which was interpreted as less severe damage than in the earlier-onset group of patients. While this might be true, the study highlights another potentially important confound; that is, the age of the control group. The mean age of controls for the contrast with "late-onset" AD was 73 years but at this age one would expect around one third of a control population to be amyloid positive, i.e., in the pre-clinical stage of AD (Villemagne et al., 2011). In contrast, the control group for the early-onset AD cohort had a mean age of 59 years—an age at which one would expect negligible, if any, contamination from pre-clinical AD. In other words, older control subjects are probabilistically more likely to introduce undesirable variability into group comparisons, which may explain some of the differences in sensitivity across studies. This suggests that when targeting "late-onset" AD, it would be desirable in future that the control group was defined by being amyloid-negative, in addition to being cognitively intact, to ensure that this possible confound was not attenuating sensitivity.

Studies in symptomatic, young PSEN1 carriers (Ryan et al., 2013), or in relatively young sporadic AD cohorts (age: 63 ± 5, Mahoney et al.), where control groups were young and abnormalities, very extensive—with posterior predilection but spreading toward anterior areas—also support the notion that control age might be a contributing factor affecting analysis sensitivity. However, extensive abnormalities are not a unique feature of young cohorts, a number of studies in older groups have also shown a similar behavior, e.g., mild AD in Fieremans et al. (2013), mild AD in Acosta-Cabronero et al. (2012), mild AD in Douaud et al. (2011), mild AD (ADNI data) in Rowley et al. (2013) and mild AD (ADNI data) in Nir et al. (2013). In such studies, the posterior parietal and superior temporal regions highlighted in **Figure 9** were key features, but abnormalities also spread more anteriorly along association tracts and also included extensively the anterior corpus callosum, corticocortical association fibers running through the external capsule and the anterior limb of the internal capsule. Disentangling whether more, or less, widespread distributions across studies are the result of subtle technical factors, differences in disease severity, in control variability, or in cohort size, is impossible to disambiguate, but it should be highlighted that from the selected studies, cohorts with the most widespread abnormality patterns had, in general, more advanced disease stages than the cohort displayed in **Figure 9**, suggesting disease severity is a dominant factor in result sensitivity. A recent study in two age-matched AD cohorts with different disease severity that controlled for all the above factors using the same acquisition, the same control cohort and the same number of subjects in each patient group, showed evidence indicating a posterior first, then more anterior distribution in more advanced disease stages (Acosta-Cabronero et al., 2012)—matching the expected progression of degeneration as would be measured by cerebral glucose metabolism (Minoshima et al., 1997; Nestor et al., 2003b). In summary, although a number of additional factors might have contributed to variability in result sensitivity across studies, disease staging appears to drive the overall distribution of WM abnormalities in AD.

Focusing on the earliest DTI changes in AD specifically, Molinuevo et al. (2014) studied a combined cohort of cognitively normal subjects and SCI with a CSF profile consistent with AD, while Acosta-Cabronero et al. (2012) reported on patients with MCI-stage AD confirmed through longitudinal follow-up. The two studies showed a consistent picture of the affected neural network in incipient stages of AD. Both results highlighted involvement of parietal and superior temporal WM bilaterally spatial distributions that are also in close agreement with the expected landscape of WM involvement that can be inferred from the pattern of glucose hypometabolism in very mild AD (Nestor et al., 2003b, 2006).

#### **DIFFERENTIAL DIFFUSION METRIC SENSITIVITY DURING THE COURSE OF AD**

An interesting by-product of TBSS analyses in AD has been in showing that alterations in diffusion behavior are complex while early (and some recent) studies tended to focus only on FA—often leading to largely insensitive results (Damoiseaux et al., 2009; Zarei et al., 2009; Balachandar et al., 2014)—, more extensive abnormality distributions for MD have been a consistent finding across studies (Acosta-Cabronero et al., 2010, 2012; Douaud et al., 2011; Huang et al., 2012; Fieremans et al., 2013; Mahoney et al., 2013; Nir et al., 2013; Rowley et al., 2013; Ryan et al., 2013). What this behavior represents in neurobiological terms, however, is unknown because, as discussed already, DTI offers limited information about the geometry of the underlying damaged microstructure. In contrast, what DTI enables us to infer from such behavior is that WM degeneration in AD results in alterations to WM tissue that are not grossly driven along any preferential direction (Acosta-Cabronero et al., 2010). Subtle orientation dependent behaviors missed by FA can be explored by dissecting the apparent diffusion coefficient encapsulated in MD into its primary components, i.e., λ<sup>1</sup> (also commonly referred to in the literature as AxD or DA) and RD. A superficial reading of studies examining these primary components appear not to provide a clear picture—a number of studies such as those in very mild AD by Acosta-Cabronero et al. (2012) and in CSFpositive SCI by Molinuevo et al. (2014) indicate that λ<sup>1</sup> might be the dominant source of DTI abnormalities; whereas other studies such as those in mild AD by Mahoney et al. (2013) and Bosch et al. (2012), and in moderately impaired AD by Ryan et al. (2013) show greater RD effects. Douaud et al. (2011) proposed that the λ<sup>1</sup> increase could be explained as an artifact of loss of one fiber population in areas of crossing fibers (see **Figure 3**). This would be a plausible theory if increased λ<sup>1</sup> were restricted to areas of crossing fibers (e.g., centrum semiovale), however increased λ<sup>1</sup> in AD is not limited to such regions: the same phenomenon is observed in the midline corpus callosum and fornix where crossing tracts are not present. This indicates that increased λ<sup>1</sup> in AD must be due to another mechanism but what that might be in biological terms is, at present, unclear2 .

The patient cohorts in Mahoney et al. (2013); Bosch et al. (2012) and the symptomatic cohort in Ryan et al. (2013), which all reported increased RD, had more advanced dementia (MMSE < 23) than those outlined with extensive λ<sup>1</sup> changes, suggesting that on closer inspection of the literature, individual component metrics of the diffusion tensor may evolve along different trajectories with disease progression. This implies that different metrics may be more or less sensitive to different disease stages. This hypothesis was tested on two—very mildly (MCI-stage) and mildly impaired—AD groups using the same MR acquisition and post-processing methods (Acosta-Cabronero et al., 2012). The distribution of increased λ<sup>1</sup> involving posterior temporo-parietal WM—largely consistent with Molinuevo et al. (2014)—was identified as the first sign of change. In the more advanced group, however, increased RD—which, in turn, drove FA reductions—emerged in these regions that had shown increased λ<sup>1</sup> in the very mild group. As degeneration spread to new areas—notably frontal and left temporal WM—increased λ<sup>1</sup>

<sup>2</sup>One possible situation where loss of crossing fibers may explain increased λ<sup>1</sup> was a study that examined the neural substrate of topographical memory impairment in AD (Pengas et al., 2012). The study in question found correlations between increasing λ<sup>1</sup> and decreasing task performance; interestingly the regions that correlated were all in close proximity to the cortical ribbon in other words, in the U-fiber layer. It is tempting, therefore, to speculate that local degeneration of GM, in turn meaning less axons entering adjacent WM, might lead to a simplification of the adjacent U-fiber layer—i.e., less crossing fibers—making increased λ<sup>1</sup> correlate significantly.

was again the first abnormality to appear. These results suggested that while λ<sup>1</sup> increase is the first sign of change in a given location, RD is relatively insensitive at the earliest disease stages but becomes progressively more abnormal as the disease progresses. Confirming this theory, a subset of AD patients in the same study, who were scanned serially, revealed that RD/FA changes emerged at follow up in areas where increased λ<sup>1</sup> had been seen at baseline. Interestingly, λ<sup>1</sup> increases did not progressively worsen over time with longitudinal follow-up. There was, furthermore, no evidence for progressive worsening of λ<sup>1</sup> between the two severity stages in the cross-sectional cohorts when investigating the midline splenium in isolation. In fact, if anything there was a slight tendency for the increase in λ<sup>1</sup> to lessen with advancing disease (Acosta-Cabronero et al., 2012); in other words suggesting that when a WM region first becomes affected in AD, there is a sharp increase in λ1, that may possibly attenuate with further disease progression.

In summary, RD progressively increases (and therefore FA progressively decreases) as a function of disease severity—as defined by degree of cognitive impairment—whereas λ<sup>1</sup> does not

**FIGURE 10 | TBSS in very mild AD.** TBSS results for *N* = 21 very mild AD patients (age: 72 ± 5, <MMSE>= 26 ± 2) vs. *N* = 26 matched controls (age: 68 ± 6) (Acosta-Cabronero et al., 2012). The patient group included *N* = 16 subjects who were scanned with a diagnosis of mild cognitive impairment but were subsequently shown to have probable AD with longitudinal follow-up. Black clusters denote increased λ<sup>1</sup> in patients at PFWE < 0.05, whereas those in blue represent changes with *post-hoc* control over the false discovery rate (FDR) at *q* < 0.05. Thresholded statistical maps are overlaid onto the TBSS skeleton and MNI152 template. Sagittal coordinates are given in millimeter (*x* < 0 is left). Abbreviations: [aTh] anterior thalamic white matter; [CGh] cingulum at the level of the hippocampus; and [FX] fornix.

progressively increase. That increasing λ<sup>1</sup> is not a function of disease severity, at least in the midline splenium, suggests that the early λ<sup>1</sup> increase in AD may be capturing an upstream event to axonal degeneration, whereas processes more directly related to neuronal loss might dominate RD/FA dynamics—implying thus that, whilst λ<sup>1</sup> in the splenium could act as a state-specific marker in prodromal disease stages, the spatial distribution of RD/FA changes may be a more suitable staging biomarker.

#### **FORNIX CHANGES IN AD**

Although atlas-based approaches consistently find strong effects in the fornix (Keihaninejad et al., 2013; Racine et al., 2014), this was not, however, a universal feature of the TBSS studies reviewed above. Such tracts are often excessively penalized due to their small size by the strictness of non-parametric statistics controlling for family-wise error (FWE) rate. Using, however, the more lenient false-discovery rate (FDR) correction (Benjamini and Hochberg, 1995) to show abnormalities that are less prominent at the cluster level—yet statistically robust—fornix changes emerge also with TBSS analysis. **Figure 10** illustrates the impact of this slight statistical modification, resulting in "new" λ<sup>1</sup> abnormalities in the fornix, parahippocampal WM and anterior thalamus (possibly capturing change in the mammillo-thalamic tract). These structures are not just random new blobs—they precisely target major WM connections in the limbic-diencephalic network that has been highlighted as the early hypometabolic landscape of AD (Nestor et al., 2003b, 2006). For a focused review on the specific role of the fornix in this network, including the possibility of it being therapeutic target for deep brain stimulation in AD, see the review in this Research Topic (Oishi and Lyketsos, 2014).

The notion that limbic tracts and the corpus callosum particularly the splenium—are key features in early stages of AD has now a solid body of evidence. The theory that these regions should show the most advanced degeneration in the course of AD (and therefore be the first to show RD/FA changes) when other areas, being less advanced in the course of disease, would only show λ<sup>1</sup> changes is strongly supported by the regional TBSS analysis in mild AD carried out by Huang et al. (2012). These authors found λ<sup>1</sup> increase in association and projection tracts and RD/FA effects in the corpus callosum, cingulum and fornix, precisely suggesting that the latter were at a more advanced stage of degeneration compared to the former. In a similarly impaired cohort extracted from the ADNI dataset, Nir et al.'s regional TBSS study also found strong RD/FA alterations for such structures. In addition, a prospective study—targeting specifically the corpus callosum, fornix and cingulum—found significant crosssectional and longitudinal tensor differences for all three structures (Keihaninejad et al., 2013). The common aspect in these studies is that—consistent with all the selected works—absolute diffusivities were increased in AD relative to controls, whereas FA was reduced.

#### **CONCLUSION**

In summary, the existing literature supports the theory that a bilateral neural network that involves preferentially the cingulum bundle, fornix and corpus callosum is vulnerable to early disease processes triggered by the Alzheimer's disease neuropathological cascade. This suggests that the nodes that this network interconnects—i.e., the mesial temporal lobe, mammillary bodies, thalamus, and posterior cingulate—degenerate as a network, rather than in isolation.

To conclude with a note of caution and a window to future research directions. While much effort was made to rationalize published DTI studies in AD for this review, it remains possible that a large number of technical factors including suboptimal acquisitions, poor tensor reconstruction routines, poor image coregistration performance, or intrinsic DTI limitations, to name a few, might still have resulted in some systematically incorrect results across the literature, in turn, leading to incompletely valid interpretations in this review. A number of technical developments, however, should hopefully soon help confirm, or reject, the above theories: stronger gradient systems enabling larger *b*values, shorter echo times and smaller voxels; improved image quality through multiband acquisitions (Frost et al., 2014); the prescription of multiple *b*-values; the implementation into standard tensor reconstruction pipelines of distortion correction, **B**-matrix rotation and multi-component fitting routines with a free-water component (Pasternak et al., 2009); the application of advanced diffusion models beyond the single tensor such as diffusion kurtosis imaging (Jensen et al., 2005), NODDI (Zhang et al., 2012), CHARMED (Assaf and Basser, 2005), or q-space approaches such as q-ball reconstruction (Tuch, 2004) or diffusion spectrum imaging (Wedeen et al., 2005); the use of *state-ofthe-art* co-registration methods (Zhang et al., 2006; Avants et al., 2008); and a long *et cetera*.

# **ACKNOWLEDGMENT**

We gratefully acknowledge the support from our patients, their relatives and healthy volunteers, without whom our past research reviewed here would have not been possible.

# **SUPPLEMENTARY MATERIAL**

The Supplementary Material for this article can be found online at: http://www.frontiersin.org/journal/10.3389/fnagi.2014. 00266/abstract

#### **REFERENCES**


Torrey, H. C. (1956). Bloch equations with diffusion terms. *Phys. Rev.* 104, 563–565. doi: 10.1103/PhysRev.104.563


between vascular dementia and Alzheimer disease. *Stroke* 40, 773–779. doi: 10.1161/STROKEAHA.108.530832


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 10 August 2014; accepted: 15 September 2014; published online: 02 October 2014.*

*Citation: Acosta-Cabronero J and Nestor PJ (2014) Diffusion tensor imaging in Alzheimer's disease: insights into the limbic-diencephalic network and methodological considerations. Front. Aging Neurosci. 6:266. doi: 10.3389/fnagi.2014.00266*

*This article was submitted to the journal Frontiers in Aging Neuroscience. Copyright © 2014 Acosta-Cabronero and Nestor. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original*

*author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Alzheimer's disease and the fornix

# **Kenichi Oishi <sup>1</sup>\* and Constantine G. Lyketsos <sup>2</sup>**

<sup>1</sup> The Russell H. Morgan Department of Radiology and Radiological Science, School of Medicine, Johns Hopkins University, Baltimore, MD, USA <sup>2</sup> Department of Psychiatry and Behavioral Sciences, Johns Hopkins Bayview and Johns Hopkins Medicine, Baltimore, MD, USA

#### **Edited by:**

P. Hemachandra Reddy, Oregen Health and Science University, USA

#### **Reviewed by:**

Mario Parra, University of Edinburgh, UK Russell H. Swerdlow, University of Kansas Medical Center, USA

#### **\*Correspondence:**

Kenichi Oishi, The Russell H. Morgan Department of Radiology and Radiological Science, School of Medicine, Johns Hopkins University, 217 Traylor Building, 720 Rutland Avenue, Baltimore, MD 21205, USA e-mail: koishi@mri.jhu.edu

Alzheimer's disease (AD) is the most common form of neurodegenerative dementia. Researchers have long been focused on the cortical pathology of AD, since the most important pathologic features are the senile plaques found in the cortex, and the neurofibrillary tangles and neuronal loss that begin in the entorhinal cortex and the hippocampus. In addition to these gray matter (GM) structures, histopathological studies indicate that the white matter (WM) is also a good target for both the early diagnosis of AD and for monitoring disease progression. The fornix is a WM bundle that constitutes a core element of the limbic circuits, and is one of the most important anatomical structures related to memory. Functional and anatomical features of the fornix have naturally captured researchers' attention as possible diagnostic and prognostic markers of AD. Indeed, neurodegeneration of the fornix has been histologically observed in AD, and growing evidence indicates that the alterations seen in the fornix are potentially a good marker to predict future conversion from mild cognitive impairment (MCI) to AD, and even from cognitively normal individuals to AD. The degree of alteration is correlated with the degree of memory impairment, indicating the potential for the use of the fornix as a functional marker. Moreover, there have been attempts to stimulate the fornix using deep brain stimulation (DBS) to augment cognitive function in AD, and ongoing research has suggested positive effects of DBS on brain glucose metabolism in AD patients. On the other hand, disease specificity for fornix degeneration, methodologies to evaluate fornix degeneration, and the clinical significance of the fornix DBS, especially for the long-term impact on the quality of life, are mostly unknown and need to be elucidated.

**Keywords: Alzheimer's disease, mild cognitive impairment, diffusion tensor imaging, deep brain stimulation, fornix, limbic system, magnetic resonance imaging, white matter**

#### **INTRODUCTION**

Alzheimer's disease (AD) research has targeted beta-amyloid and tau pathologies as diagnostic or progression markers. White matter (WM) shows less pathologic evidence of betaamyloid or tau protein; nonetheless, alterations are seen in WM as well. Among WM bundles reportedly abnormal in AD, the fornix is especially interesting because robust and consistent alterations have been demonstrated. These could predict future gray matter (GM) atrophy and conversion from mild cognitive impairment (MCI) to AD, and even from cognitively normalcy to AD. Moreover, there are ongoing efforts to impact the biologic processes of AD by stimulating the fornix using deep brain stimulation (DBS). This review focuses on the fornix as a potential alternative diagnostic and prognostic marker, as well as a potential therapeutic target in AD.

#### **WM ALTERATIONS SEEN IN AD**

Histopathological studies have revealed various types of WM alterations in AD. Symmetrical WM changes compatible with incomplete infarction, independent of the GM changes, have been reported. These WM changes are partly explained by co-existing small vessel disease (Brun and Englund, 1986; Englund et al., 1988). WM changes are also related to cerebral amyloid angiopathy of AD, supporting a vascular contribution to the WM alterations (Haglund and Englund, 2002). Chronic hypoxia caused by vascular alterations might induce oxidative stress that could damage axonal membranes and myelin, and the reactivity of oligodendroglia to such damage (Back et al., 2011). However, in preclinical AD, WM alterations are independent of vascular alterations (de la Monte, 1989). In addition, degeneration of the WM was more obvious than that of the GM in early stages (de la Monte, 1989), suggesting that WM alteration precedes GM pathology. Indeed, the earliest disease-related anatomical changes have been observed in the axons and dendrites of AD animal model (Gunawardena and Goldstein, 2001; Pigino et al., 2003; Stokin et al., 2005; Chevalier-Larsen and Holzbaur, 2006). Therefore, mechanisms that cause the early pathological changes, aside from vascular alterations, might exist. According to the updated amyloid cascade hypothesis (Hardy and Selkoe, 2002), one of the high upstream events is production of soluble betaamyloid oligomers, been discovered to be toxic to WM structure and function (Lue et al., 1999; Xu et al., 2001; Lee et al., 2004). This could be a major cause of the primary WM alterations in AD (Roher et al., 2002; Chalmers et al., 2005) at least in early stages.

In addition to widespread WM alterations, structure-specific WM alterations have been observed, especially in limbic structures (Damoiseaux et al., 2009; Acosta-Cabronero et al., 2010; Smith et al., 2010; Liu et al., 2011). This might at least partly be explained by aberrant axonal transport due to misregulation of tau seen in early disease stage (Alonso et al., 1994; Ebneth et al., 1998; Dawson et al., 2010), with resultant tau aggregation that is closely related to neurodegeneration (Braak and Braak, 1995). Since the neurodegeneration of AD usually propagates systematically from the transentorhinal area to limbic structures, to neocortex at the most advanced stage, associated Wallerian-like degeneration (Bossy-Wetzel et al., 2004; Coleman, 2005) likely follows the same order, which might explain structure-specific WM alterations in AD (Bramblett et al., 1992).

Congruent with histopathological findings, WM alteration has been observed *in vivo* using various imaging modalities. Among these, diffusion tensor imaging (DTI) is one of the most effective modalities for the investigation of the WM anatomy since DTI has the ability to visualize WM fibers, based on directionality, and has the capability to provide detailed information about microscopic organization of the fibers. A meta-analysis of DTI studies indicated that WM alterations in AD are widespread throughout the brain (Sexton et al., 2011), particularly in limbic fibers, direct connections to the medial temporal lobe and the most vulnerable WM structures (Rose et al., 2000; Kantarci et al., 2001; Medina et al., 2006; Ringman et al., 2007; Stahl et al., 2007; Zhang et al., 2007; Zhou et al., 2008; Damoiseaux et al., 2009; Mielke et al., 2009; Salat et al., 2010). WM damage quantified by DTI has been correlated with atrophy in anatomically connected GM areas in AD patients, but the correlation was not robust in patients with amnestic MCI (Agosta et al., 2011). These findings support the histopathological observation that primary WM alterations may precede both GM degeneration, as well as secondary degenerative processes of the WM after neuronal loss. Based on the evidence about WM pathology in AD, Sachdev et al. proposed incorporating WM alterations into the model (Jack et al., 2010) of the pathophysiological cascade in AD (Sachdev et al., 2013).

#### **WHY IS THE FORNIX IMPORTANT?**

The fornices are WM bundles that originate from the bilateral hippocampi, merge at the midline of the brain, again divide into the left and right side, then into precommissural and postcommissural fibers, terminating at the septal nuclei, nucleus accumbens (precommissural fornix), and hypothalamus (postcommissural fornix). The fornix constitutes a core element of the limbic circuit that is vulnerable in AD, and one of the most important anatomical structures related to episodic memory, impairment of which is an initial symptom of AD (Hopper and Vogel, 1976; Mehraein and Rothemund, 1976). The fornix is also related to the cholinergic dysfunction characteristic of AD (Milner and Amaral, 1984; Wenk et al., 1987; Ransmayr et al., 1989; Sara, 1989; Schegg et al., 1989; Bunce et al., 2003; Colom et al., 2010). These functional and anatomical features have naturally captured researcher attention seeking diagnostic and prognostic markers of AD. Indeed, alterations of the fornix, such as demyelination or axonal loss, were reported as early as in 1976 (Hopper and Vogel, 1976) and have been consistently identified in AD (Ringman et al., 2007; Teipel et al., 2007; Mielke et al., 2009; Acosta-Cabronero et al., 2010; Salat et al., 2010; Liu et al., 2011; Douaud et al., 2013).

The other important anatomical feature of the fornix, compared to other WM areas, is its location. The body of the fornix is readily identifiable in the mid-sagittal plane of a brain MRI, without other WM structures adjacent to it. For research using brain MRI, this feature is particularly important. The thickness of the fornix can be easily measured on structural MRI, using T1 and T2-weighted images. Generally, identification of the boundaries of WM structures on structural MRI is not easy because the boundaries between adjacent WM structures are often invisible: the fornix is one of several exceptions. DTI is a good choice with which to identify the boundaries of WM bundles. Although a drawback of DTI is susceptibility to B0 distortion, especially for areas close to air cavities (e.g., nasal and oral cavities, frontal, ethmoid, sphenoid, and maxillary sinuses, and mastoid antrum), the fornix is less affected because there is a sufficient distance from the fornix to these air cavities. The anatomical location is also attractive from a neurosurgical point of view, since the fornix is surgically accessible for implanting electrodes for DBS (see Section Treatment), while other limbic structures are not easily accessible.

# **IMAGING STUDY OF THE FORNIX IN AD REGION-OF-INTEREST STUDIES**

Atrophy of the fornix measured using T1-weighted MRI has often been reported in AD (Callen et al., 2001; Copenhaver et al., 2006), although the onset of this atrophy has been a subject of debate. Fletcher et al. (2013) reported that atrophy of the fornix is a strong predictor of conversion from normal cognition to MCI or AD, suggesting that atrophy is present in the fornix at an early disease stage, before clinical manifestations. Abnormalities in the fornix have also been identified using DTI. A reduction of fractional anisotropy (FA) and an increase in diffusivity in the fornix is a robust and consistent finding in AD as demonstrated with manual ROI- analysis (Ringman et al., 2007; Mielke et al., 2009; Oishi et al., 2012), and the extent of the changes in DTI-derived parameters have been correlated with cognitive decline (Mielke et al., 2009; Oishi et al., 2012). Notably, FA reduction was already present in an asymptomatic gene carrier of familial AD who went on to develop AD in the future (Ringman et al., 2007). Longitudinal observation of cognitively normal elderly or individuals with amnestic MCI indicated that reduced fornix FA predicts conversion from normal cognition to amnestic MCI and from amnestic MCI to AD (Oishi et al., 2012). Moreover, lower fornix FA predicted later cognitive decline and hippocampal atrophy (Mielke et al., 2012). These studies suggest that volume loss and FA reduction of the fornix is one of the earliest anatomical changes in AD that happens before clinical manifestations.

#### **WHOLE-BRAIN STUDIES**

While region-of-interest analysis is a good choice when there are specific hypotheses (e.g., the fornix is affected in AD), the drawback is the hypothesis dependency. To evaluate the spatial specificity of these findings, whole-brain analysis is a good choice. Voxel-based analysis is one of the most widely used approaches for whole-brain analysis, and has been used to evaluate regional abnormalities in WM volume (Li et al., 2008; Balthazar et al., 2009; Guo et al., 2010; Serra et al., 2010; Yoon et al., 2011) or in DTI-derived parameters, such as anisotropy and diffusivity (Head et al., 2004; Medina et al., 2006; Xie et al., 2006; Teipel et al., 2007; Zhang et al., 2009). Although cross-sectional group comparisons have identified widespread WM abnormalities in AD, the above studies failed to identify reduced volume or FA in the fornix, except for a study that used multivariate analysis (Teipel et al., 2007). The main explanation is probably inaccuracy in image normalization: the accuracy of automated non-linear registration used in these studies was limited for thin structures like the fornix, in which only a few pixels of misregistration cause significant loss of sensitivity to detect differences between groups (Oishi et al., 2009). A common approach to solve this problem is to apply Tract-Based Spatial Statistics (TBSS; Smith et al., 2006), in which WM tracts are "skeletonized" to summarize the anatomy of WM structures, from which statistics are calculated. Since the scalar values perpendicular to the skeleton are projected onto the skeleton, the effect of WM misregistration is largely ameliorated. Most of the whole brain TBSS studies applied to AD have, indeed, identified an FA reduction in the fornix, as well as in other limbic fibers, even in earlysymptomatic patients or in individuals at high risk for developing AD (Damoiseaux et al., 2009; Honea et al., 2009; Stricker et al., 2009; Zarei et al., 2009; Acosta-Cabronero et al., 2010; Bosch et al., 2010; Smith et al., 2010; Liu et al., 2011; Douaud et al., 2013).

Application of large deformation diffeomorphic metric mapping (LDDMM) to DTI is another way to increase the sensitivity of whole-brain analysis, by which registration accuracy of the WM bundles could be increased. By using LDDMM, significant volume loss, FA reduction, and increases in mean diffusivity (MD) and radial diffusivity were detected in the fornix of individuals with AD (Oishi et al., 2011a,b). **Figure 1** is an example of such an analysis, designed to find brain areas with AD-specific WM alterations. This analysis indicates significant FA reduction in the fornix, the splenium of the corpus callosum, as well as in several small areas in superficial WM in the frontal lobes. To further increase sensitivity, multivariate models (Ashburner and Kloppel, 2011), such as principal component analysis (PCA; Teipel et al., 2007), or to apply voxel-grouping methods, such as atlas-based analysis (ABA; Oishi et al., 2009), were applied: all of these detected abnormalities in the fornix in AD. These whole-brain studies using sophisticated approaches consistently indicate the fornix as the "hotspot", in which the earliest anatomical changes begin ("Fornix First" (Landhuis, 2013)), and where the most robust changes are seen in AD.

#### **PITFALLS AND ISSUES IN FORNIX IMAGE ANALYSIS Misregistration**

As mentioned previously, TBSS is currently a standard method by which to ameliorate WM tract misregistration. However, in patients with ventricular enlargement, like AD, the fornix is strongly deformed by the elevated and stretched corpus callosum, which results in uncorrectable misregistration, even with TBSS (Hattori et al., 2012). Therefore, careful interpretation is needed for AD studies. An inverse-projection of TBSS results onto original images is a simple method with which to identify invalid results caused by misregistration.

One solution to reduce the impact of erroneous mapping of voxels, and the consequent mis-parcellation of target structures, is to adopt multi-atlas approaches. Rather than using a single atlas for anatomical labeling, this approach uses multiple atlases to compute the likelihoods of labels, which are then fused to create an anatomical parcellation map for each image. This approach, which accounts for variations in brain anatomy, has been adapted to parcellate WM structures in DTI (Tang et al., 2014) with high accuracy even for individuals with ventricular enlargement. Applicability of the multi-atlas label fusion method to accurately define the fornix is a promising approach, which needs to be developed in the future with automated image analysis of AD.

# **Contamination of signal from cerebrospinal fluid**

The fornix is prone to partial volume effects caused by contamination of signal from the cerebrospinal fluid (CSF), since it is a thin bundle that runs through the lateral ventricles that are filled with CSF. The effect is especially seen in individuals with AD because of atrophy, and thus, increased CSF signal in each voxel (Pfefferbaum and Sullivan, 2003; Metzler-Baddeley et al., 2012; Berlot et al., 2014). The greater diffusivity in CSF relative to the WM structures leads to false elevation of diffusivity values and a false reduction in FA values (Alexander et al., 2001; Vos et al., 2011). Therefore, reduced FA or increased diffusivity identified in the DTI studies of AD probably reflects pathological alterations of the fornix itself plus the partial volume effects caused by atrophy. Although this dual effect (atrophy + microstructural change) on the DTI parameters might make DTI a sensitive tool with which to detect early anatomical changes related to AD, it complicates inferences about the underlying anatomical changes. There have been several attempts to ameliorate the effects of CSF contamination in order to draw inferences about the microstructural changes in the WM structures, such as CSF-suppressed diffusion imaging (Kwong et al., 1991), or the free water elimination and mapping method that is applicable at the post-acquisition stage (Pasternak et al., 2009), both of which could detect alteration of the fornix in AD (Kantarci et al., 2010; Fletcher et al., 2014). Investigations of underlying changes in WM structures seem to suggest a good indication for the use these advanced methods (Metzler-Baddeley et al., 2012).

#### **Cause of fornix degeneration and DTI parameters of choice**

Since the contribution of each factor (see WM alterations seen in AD**)** to the degeneration of the fornix might vary depending on the degree of disease progression, longitudinal observation of the histopathology, from the preclinical to the advanced stages, is especially important. Use of the AD model in animals is an option for such longitudinal observation. A reduced size of the fornix is seen in the PDAPP mice (Gonzalez-Lima et al., 2001) and transgenic APP/PS1 mice (Delatour et al., 2006), and

**FIGURE 1 | Voxel-based group comparison between 19 AD and 22 cognitively normal age-matched control participants (NC), indicating the fornix as the hotspot**. Areas with signal or volume alterations in AD compared to NC are shown as colored maps, overlaid on an averaged FA map **(A–E)**, an averaged T<sup>2</sup> map **(F)**, and an averaged GM segmentation map **(G)**. **(A)** Areas with reduced FA. **(B)** Areas with increased MD. **(C)** Areas with increased λq. **(D)** Areas with increased λ⊥. **(E)** Areas with an increased and decreased Jacobian, which was calculated from a transformation matrix obtained from the normalization of DTI. **(F)** Area with increased T2. **(G)** Areas with increased and decreased Jacobian, which were calculated from a

transformation matrix obtained from the normalization of a GM segmentation map. Pink arrows with numbers indicate: 1 = the fornix; 2 = the cingulum; 3 = the posterior cingulate gyrus white matter; 4 = splenium of the corpus callosum; 5 = genu of the corpus callosum; 6 = the prefrontal white matter; 7 = the orbitofrontal white matter; the temporal white matter; 9 = the parietal white matter; 10 = the periventricular area; 11 = the thalamus; 12 = the hippocampus; 13 = the entorhinal area; 14 = the parietal cortex; 15 = the area between the caudate head and the gyrus rectus; 16 = the lateral ventricle. White arrows show the misregistration seen in the left posterior horn of the lateral ventricle (From Oishi et al., 2011a; with permission).

immunohistochemistry shows the presence of increasing amyloidosis and related microgliosis and astrogliosis (Maheswaran et al., 2009). However, longitudinal changes in the size of the fornix differ depending on the type of AD model mice. Since type and variations in gene mutation are relevant to the pathophysiology of AD model animals, and considerable differences are seen between human AD and AD model animals, human *in vivo* neuroimaging studies are important to understand the pathophysiological underpinnings. For early AD, diffusion measures (MD and axial and radial diffusivities) could detect the widespread pathology

**FIGURE 2 | Example of the fornix sign**. The axial (left), coronal (middle), and sagittal (right) slices of the color-scaled FA map are shown with the magnified view of the fornix (yellow rectangle). **(A)** A cognitively normal 81-year-old woman without the fornix sign. The core part of the fornix appears yellow to red (FA 0.5–0.8). **(B)** A cognitively normal 80-year-old man with the fornix sign. The fornix appears green (FA < 0.5). He converted to amnestic MCI (aMCI)

1 year after the scan. **(C)** An 80-year-old man with aMCI without the fornix sign. He was stable during 3 years observation period after the scan. **(D)** A 79-year-old man with aMCI with the fornix sign. He converted to AD 1 year after the scan. **(E)** A 74-year-old man with aMCI without the fornix sign. He converted to dementia with Lewy bodies 3 years after the scan. **(F)** An 81-year-old woman with AD with the fornix sign. FA, fractional anisotropy.

of AD more sensitively than FA (Acosta-Cabronero et al., 2010; Oishi et al., 2011a). However, for the fornix, there seems to be a tendency that a decrease in FA is more sensitive than an increase in diffusivity measures. Whether there is a difference in pathophysiology between the fornix and other WM areas remains to be elucidated.

# **CLINICAL APPLICATIONS OF THE FINDINGS SEEN IN THE FORNIX**

#### **DIAGNOSIS AND PREDICTION**

There are several hurdles when applying research findings to clinical practice. Research studies are based on selected participants with strict inclusion and exclusion criteria, and usually age- and gender-matched participants are recruited as controls. However, routine image-based diagnosis is almost exclusively performed by qualitative visual inspection, which is applied to each individual for clinical decision-making. Although numerous publications link AD to degeneration in the fornix, volume or the diffusivity are not quantified in daily practice because of the time-consuming nature of this process, and the inconsistent quantification methods used in research. The fornix sign, a discoloration of the fornix observed on color-coded FA maps of AD (Oishi et al., 2012), is one attempt to define qualitative image signs applicable to clinical images (**Figure 2**). Although sensitivity was limited, specificity was high. The positive likelihood ratio differentiating AD from controls and hence the ability to predict conversion of cognitively normal individuals to amnestic MCI, or from amnestic MCI to AD, was strong. However, further efforts are needed to identify appropriate scan parameters, cut-off values for the diagnosis and prediction, as well as to investigate degeneration in the fornix seen in other types of dementia.

Given that individuals with different types of memory complaints visit clinics, and physicians need to identify the etiology of the memory complaint for medical decision-making, the value of imaging signs that can separate AD from individuals with normal cognition are limited. Rather, clinically relevant imaging signs must provide clues that can be used to separate AD from other diseases that cause memory impairment. Indeed, the fornix might be affected by tumors, infections, inflammations (Yamamoto et al., 1990), metabolic abnormalities, vascular diseases (Molino et al., 2014), malformations, trauma, and hydrocephalus (Hattori et al., 2012) (for review, Thomas et al., 2011). Psychiatric diseases, such as schizophrenia and bipolar affective disorder (Oertel-Knöchel et al., 2014), or medical conditions, such as heart failure (Wu et al., 2007), also cause damage in the fornix. Less is known about the involvement of the fornix in other neurodegenerative dementias, such as frontotemporal dementia (FTD), dementia with Lewy bodies (DLB), and corticobasal degeneration (CBD). A study indicated substantial FA reduction in the fornix of the behavioral variant FTD that correlated with memory decline (Hornberger et al., 2012), and other studies have indicated that the fornix is affected in Pick's disease, FTD with tau mutation, or TAR-DNA-binding protein-43 type C pathology, but not for FTD caused by a progranulin mutation or with fused-in-sarcoma protein accumulation (Rohrer et al., 2011; Hornberger et al., 2012). Changes in the fornix were not detected by group comparison between DLB and control groups (Kantarci et al., 2010), or between CBD and control groups (Rohrer et al., 2011), but other research has reported reduced FA in the fornix of some DLB patients, especially for those with hippocampal atrophy (Firbank et al., 2011). The appropriate clinical use of imaging modalities for the evaluation of the fornix is, therefore, an important future direction.

#### **TREATMENT**

Inspired by neuroimaging evidence about early impairment of the limbic circuit seen in AD, and an unexpected augmentation of recollection during the hypothalamic/fornix DBS performed on a patient with obesity (Hamani et al., 2008), DBS of the fornix has been proposed as a novel therapy for AD to improve the neuronal circuitry involved in memory (Laxton et al., 2010). Based on the hypothesis that an important aspect of AD is that it is a system-level disorder affecting several integrated pathways related to memory and cognition (Lyketsos et al., 2012), and supported by rodent studies indicating that DBS improved memory and produced hippocampal neurogenesis (Toda et al., 2008; Hamani et al., 2011), a clinical trial evaluating the effect of DBS targeting the fornix is ongoing. Electrodes are placed adjacent to the fornix for high frequency electronic stimulation, with the hope of an immediate effect (direct effects of stimulating the fornix) and a disease-modifying effect (neurogenesis). Although its number of participants was small, the earlier Phase 1 trial suggested improved glucose metabolism in the posterior cortical areas, especially in cognitively improved participants after fornix DBS (Laxton et al., 2010; Smith et al., 2012). The assessment of long-term impact on cognitive function, activities of daily living, quality of life, and mortality is an important future direction.

# **ACKNOWLEDGMENTS**

The authors thank Ms. Mary McAllister for help with manuscript editing. This study was supported by NIH grants, R01AG031348, R01AG042165 (Constantine G. Lyketsos), R21AG033774, P50AG 005146, and R01HD065955 (Kenichi Oishi). Its contents are solely the responsibility of the authors and do not necessarily represent the official view of NIA, NICHD, or NIH.

# **REFERENCES**


**Conflict of Interest Statement:** Constantine G. Lyketsos has received support from the following organizations: Associated Jewish Federation of Baltimore, Weinberg Foundation, Forest, Glaxo-Smith-Kline, Eisai, Pfizer, Astra-Zeneca, Lilly, Ortho-McNeil, Bristol-Myers, and Novartis. Constantine G. Lyketsos has served as a consultant/advisor for Astra-Zeneca, Glaxo-Smith-Kline, Eisai, Novartis, Forest, Supernus, Adlyfe, Takeda, Wyeth, Lundbeck, Merz, Lilly, and Genentech. Constantine G. Lyketsos has received honorarium or travel support from Pfizer, Forest, Glaxo-Smith-Kline, and Health Monitor. This arrangement has been approved by the Johns Hopkins University in accordance with its conflict of interest policies.

#### *Received: 02 July 2014; accepted: 22 August 2014; published online: 11 September 2014*. *Citation: Oishi K and Lyketsos CG (2014) Alzheimer's disease and the fornix. Front. Aging Neurosci. 6:241. doi: 10.3389/fnagi.2014.00241*

*This article was submitted to the journal Frontiers in Aging Neuroscience*.

*Copyright © 2014 Oishi and Lyketsos. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms*.