Edited by: Daniel Gardner, Weill Cornell Medical College, USA
Reviewed by: Jan G. Bjaalie, University of Oslo, Norway; Badri Roysam, Rensselaer Polytechnic Institute, USA; Alvaro Duque, Yale University School of Medicine, USA
*Correspondence: Slawomir J. Nasuto, School of Systems Engineering, University of Reading, Whiteknights, Reading, Berkshire RG6 6AY, UK. e-mail:
This is an open-access article distributed under the terms of the
The ability to create accurate geometric models of neuronal morphology is important for understanding the role of shape in information processing. Despite a significant amount of research on automating neuron reconstructions from image stacks obtained via microscopy, in practice most data are still collected manually. This paper describes Neuromantic, an open source system for three dimensional digital tracing of neurites. Neuromantic reconstructions are comparable in quality to those of existing commercial and freeware systems while balancing speed and accuracy of manual reconstruction. The combination of semi-automatic tracing, intuitive editing, and ability of visualizing large image stacks on standard computing platforms provides a versatile tool that can help address the reconstructions availability bottleneck. Practical considerations for reducing the computational time and space requirements of the extended algorithm are also discussed.
Dendritic and axonal morphology plays an important role in determining neuronal behavior in health (van Elburg and van Ooyen,
Although two-photon microscopy (Denk et al.,
The development of such techniques and increasing computational power and memory allow the collection of greater amounts of morphological data and execution of more complex analyses. The purpose of semi-automatic methods is to provide significant assistance in tracing neurites; rather than forcing the user to manually segment each point along a dendrite, clicking on two positions on a neurite will automatically trace along it. Both Imaris FilamentTracer and the freeware NeuronJ perform semi-automatic tracing through the application of steerable Gaussian filters (Freeman and Adelson,
Theoretically, fully automatic tracing should be able to produce a full and accurate 3D reconstruction of a neuron from an image stack with minimal user-input. Hence, in principle, fully automatic methods should be highly preferable to semi-manual tracing. In practice, however, most tracing is still performed semi-manually with applications such as Neurolucida. The primary reason for this is inaccuracy: the time required to edit the results of an automatic reconstruction in order to obtain the desired accuracy is greater than the time required to perform a semi-manual reconstruction. Additionally, such algorithms tend to be restricted to high-quality imaging technologies such as confocal or electron microscopy (Rodriguez et al.,
A large collection of neuronal reconstructions is freely available at
In order to increase neuronal reconstruction throughputs, software development needs to address the main stages of the process: automating tracing, editing, and visualizing reconstructions. The need for increasing automation has motivated the recent DIgital reconstruction of Axonal and DEndritic Morphology (DIADEM) Challenge and the resulting competition aimed at stimulating advancement of automated morphology reconstruction software
Neuromantic
Neuromantic is a stand-alone freeware application programmed in Borland C++ Builder; it is designed to provide a simple and intuitive interface for the exploration of serial image stacks and the reconstruction of dendritic trees. Once a stack of images is loaded (JPEG, BMP, and single/multi-page TIFF file are all supported), it can be explored effectively to translate, scale, and move through the data using the mouse via a simple click-and-drag interface. The morphology may also be easily modified by deleting segments/branches or changing connectivity to correct errors.
Reconstructions are stored in the freeware SWC format (Cannon et al.,
In order to add a segment to the current reconstruction, a line is dragged orthogonally across a dendrite from edge-to-edge, thus providing an estimate of the diameter of the dendrite at that point. The parent of subsequent segments is then set to the most recently added one. Once a given dendrite has been completed, a previous branch point may be selected by left-clicking, and then subsequent segments will follow on from there.
The current slice in the stack, and thus the
There are also several modes available for overlaying the current reconstruction over the stack. It may be displayed as a simple skeleton or a series of varying width rectangles to illustrate each segment’s radius. Also, the segments themselves may be colored according to either their type or their distance from the currently viewed image slice. Finally, there is an option to hide segments that are not near the current plane of focus, thus helping to avoid visual clutter during segmentation.
Figure
Neuromantic also includes some useful real-time image processing options to aid reconstruction. With TLB stacks, where the neurites are dark on a light background, the luminosity may be inverted to allow more details to be observed in the neurites; the contrast may also be adjusted as desired through histogram stretching. These changes are only performed when drawing the visible image and do not affect the underlying stack data, thus preventing information loss.
The semi-automatic reconstruction capabilities of Neuromantic are based on the LiveWire algorithm (Barrett and Mortensen,
The algorithm uses Steerable Gaussian Filters (Freeman and Adelson,
Subsequently, the actual neurite path is calculated by applying Dijkstra’s graph optimization algorithm (Dijkstra,
The algorithm may be readily extended to 3D (as previously implemented in other LiveWire variations for segmentation) by taking into account the new image geometry when expanding nodes. The image processing remains essentially the same, except that every image in the stack is now processed in the same way to estimate neuriteness and vector flow.
The primary function of the image processing is to score pixels based on their likelihood that they belong to a neurite, as well as producing an estimate of the direction of each neurite for a given pixel. For the neurite tracing application, it is important for image processing to be computationally efficient, as the system needs to update fast enough for real-time user interaction. Steerable filters were therefore employed (Meijering et al.,
Convolving the image
where
However, following the example of Meijering et al. (
This modified Hessian implicitly represents a more elongated version of the steerable filter. The eigenvalues and corresponding eigenvectors of the modified Hessian
The eigenvalues and eigenvectors of the modified Hessian can be calculated from those of the standard Hessian:
The measure of neuriteness, ρ(
where λ(
The normalizing term of λmax in (4) is the greatest absolute eigenvalue
The original equation (
In both definitions (4) and (6), the neuriteness value is bounded such that ρ(
The primary parameter affecting the estimation quality of ρ(
Conversely, a high value of σ will lead to poor tracking on thin dendrites as some curvature will be lost by the Gaussian smoothing (fourth panel of Figure
The directional flow of the dendrite,
The Dijkstra algorithm (Dijkstra,
The algorithm employs two lists, the
Take the node
Add to the
Repeat steps 1 and 2 until the desired destination node is added to the
In this way, once a given node has been considered, the optimal cost path from the source to that node is immediately known. Another useful property is that if a series of nodes
Extending the Dijkstra algorithm from two to three dimensions is straightforward: instead of adding just the 8 pixel-neighborhood of a pixel
The cost function is fundamental to the neurite tracing algorithm, as it determines which overall route will be optimal. In general, the cost should be inversely related to the “likelihood” that the given pixel belongs to a neurite.
Let the vector
The first of the cost function terms used in Dijkstra algorithm is the
Meijering’s cost function is a linear combination of these two terms, with a weighting parameter γ, and calculates the cost of moving from pixel
The neuriteness cost,
such that the cost of the pixel is inversely proportional to its neuriteness. The second term, which penalizes the movement when the neurite flow differs from the proposed pixel-to-pixel movement, is defined as
where
and
The vector flow term,
Of the two terms, the neuriteness is considerably more important. The tracing algorithm still functions effectively when γ = 1, but completely fails as γ → 0.
Meijering’s original algorithm did not take into account normalization for diagonal pixels. The cost function for this step is multiplied by
For the expansion to 3 dimensions, changes in
Integrating
Subsequent to optimizing pixel-by-pixel routing with the Dijkstra algorithm, the final solution is obtained by sub-sampling this route. The
One minor issue with strict graph optimization is that the algorithm is generally biased toward physically shorter routes (i.e., those traversing the fewest pixels). This was observed in Meijering et al. (
The image stacks used for this type of reconstruction can easily reach one Gigabyte in size, resulting in estimated 20 GB of RAM needed for proper operation, based on required 20 bytes per pixel and 8-bit grayscale images. Thus, image processing the entire stack would be expensive in terms of both time and space.
A practical solution to this problem is to process, in real-time, smaller patches of the image stack, rather than pre-processing the entire stack. Therefore, when the user initially clicks on the 3D start point for the dendrite, a stack of patches is added that is centered on that point (usually 128 × 128 pixels), encompassing several patches above and below the current
Neuromantic allocates new patches dynamically as the user moves the cursor along the neurite; when the mouse is moved over an area not containing a patch, a new patch stack is allocated and linked in the routing algorithms so that a trace may be created across any number of patches.
The optimal solution found using patchwork is not necessarily identical to the theoretical optimum calculated without it, although in most cases they coincide. For example, for a given set of patches, after a certain amount of processing every node in those patches will have been evaluated by the Dijkstra algorithm, leaving an empty Open list. If a new patch were added after this happened, no further routing would take place as all nodes would be already analyzed.
To avoid this problem, when a new patch is added all nodes that have already been routed to at the edge of any overlapping existing patch are re-added to the Open list, such that the routing may continue onto the new patch. However, because some nodes with a greater cost than the lowest nodes in the new patch may have already been expanded, the strict guarantee of optimality is lost. In practice, this may only have detectable effects on meandering dendrites moving from one patch to another and then back again, but it has no effect if the second patch is added before routing reaches the edge of the first one, which is the usual case.
An experiment was performed to examine the semi-manual reconstruction capability of Neuromantic, the time required to complete a reconstruction, and the statistical properties of these reconstructions compared with Neuron_Morpho and Neurolucida reconstructions of the same neuron.
The trial consisted of ten participants (postgraduate student volunteers at the University of Reading’s), each of whom reconstructed the basal dendrites of a CA1 rat hippocampal neuron (as described in Section
All the participants worked with a luminosity-inverted version of the stack as the dendrite details were more apparent. They were also able to alter the contrast to highlight branches more effectively, but advised to keep it at one setting throughout the reconstruction.
Participants were given step-by-step instructions for adding new segments and branch points, and example images of how correctly segmented branches should look, enabling more effective identification, and tracing of dendrites with high spine density.
The key to routing algorithm accuracy is the quality of its cost function. The cost function should be monotonically decreasing with increasing likelihood that the pixel belongs to a neurite.
We examined the effect of applying exponential function to the cost terms on the tracing quality. Integer exponents were selected because they would help penalize areas with low neuriteness and help reduce the incidence of shortcuts taken over non-neurite pixels. Also, they are highly efficient to compute, so the speed of the algorithm would not be significantly reduced.
The considered cost function is generalized to
where
The original cost function is a specific case of the generalized function where
Two main metrics were used in order to assess reconstruction accuracy – midline tracking (considering
Let the series of 3D points representing the ground-truth be ω1…
For each of the ground-truth segment points, ω
where α ∈ [0, 1] was the proportional distance along the line. This allows the estimation of a value of the
Eight different cost functions combining different polynomial terms of the neuriteness and neurite flow were examined to investigate neurite tracing accuracy:
where
A
For comparison, the final paths were subsampled by a factor of 5.
To be recommended, a given cost function variant must perform significantly better than the standard function
The Bonferroni correction for multiple comparisons was applied (Miller,
The null hypothesis,
In the case of testing the varying values of η, each other value will be compared against a value of η = 1.0, as this represents the default case of no biasing for moving between different image slices.
The benchmark data used to evaluate manual reconstruction as well as semi-automatic tracking came from 200 μm brain sections from adult, male, Sprague-Dawley rats (Desmond et al.,
The original Neuron_Morpho and Neurolucida reconstructions from (Brown et al.,
For the semi-automatic reconstruction experiments five branches were selected as benchmarks and manually reconstructed by the first author to obtain the ground-truth against which the semi-automatic reconstruction could be assessed.
The second set of benchmarks for semi-automatic reconstruction comes from a guinea pig piriform cortex neuron labeled with Neurobiotin, (Libri et al.,
Five branches were carefully segmented using the semi-manual capabilities of Neuromantic as test cases. Analogously to Meijering et al. (
Example images from the two benchmark stacks are shown in Figure
The number of reconstructed segment per time produced by participants varied from 861 and 4549, and the overall time taken from 140 to 290 min. The segments added per minute ranged from 4.5 to 15.7, with a mean value of 10.2. For comparison, Brown et al. (
The number of segments per time is used here as an index indicative of the ease of use of reconstruction software, eliminating variations due to the average segment length. However, it is worth noting that the amount of segments in semi-manual reconstructions produced by each participant varied significantly, despite the fact that the introductory demonstration included a recommended segment size in order to reduce this issue. This is mainly attributed to the varying desire of the participants to complete the task as quickly as possible, and seems unavoidable in a trial of this kind without imposing some physical segment length limit within the software itself.
A variety of statistical measures from the reconstructions were calculated using L-Measure (Scorcioni et al.,
Figure
The most obvious gross morphological difference between the Neuron_Morpho and Neurolucida reconstructions is that the former lacks the large branch furthest to the right. All Neuromantic reconstructions contain at least part of this branch.
Significant variation in radius estimation can be clearly seen over the reconstructions: visually reconstruction 5 appears to have the thinnest dendrites (closest to the original Neurolucida reconstruction). Reconstruction 2, on the other hand, demonstrates the widest dendrites, and should therefore have the largest overall volume and surface area. These observations are confirmed by the statistical analysis.
Segment data were post-processed in order to remove obvious reconstruction errors (e.g., tracing an obvious length of axon rather than dendrite, tracing a wildly out-of-focus dendrite, initiation of the reconstructions at different points on the soma). This resulted in minor discrepancy between the numbers of segments used for calculation of speed of reconstruction (original raw numbers of segments used in Table
Neuron characteristics | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
---|---|---|---|---|---|---|---|---|---|---|
# Segments | 4599 | 1815 | 4057 | 859 | 2130 | 2125 | 1431 | 1924 | 2208 | 2033 |
# Branches | 54 | 52 | 48 | 47 | 34 | 41 | 35 | 56 | 40 | 41 |
Total length (μm) | 4944.7 | 5188.3 | 4348.3 | 4448.2 | 3543.3 | 4746.1 | 3221.0 | 4633.7 | 4131.1 | 3999.1 |
Total area (μm2) | 23952.7 | 33657.2 | 24868.6 | 26040.0 | 16777.6 | 25542.4 | 22080.1 | 21983.5 | 22400.6 | 24849.0 |
Total volume (μm3) | 14028.3 | 24463.4 | 16845.5 | 16399.7 | 9553.3 | 16480.0 | 15910.7 | 13361.6 | 14166.7 | 16858.5 |
Property | Median | LQ | UQ | NM | NL |
---|---|---|---|---|---|
# Segments | 2079 | 1623 | 3132.5 | 2259 | 2572 |
# Branches | 44 | 37.5 | 53 | 36 | 32 |
Total length (μm) | 4398.3 | 3771.2 | 4845.4 | 3618.64 | 3627.3 |
Total area (μm2) | 24400.9 | 22031.8 | 25791.2 | 19469.1 | 17321.9 |
Total volume (μm3) | 12672.2 | 10347.55 | 13605.6 | 9189.1 | 7059.8 |
Measure (mean) | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
---|---|---|---|---|---|---|---|---|---|---|
Diameter (μm) | 1.512 | 1.951 | 1.732 | 1.948 | 1.474 | 1.69 | 2.113 | 1.487 | 1.719 | 1.939 |
Path distance (μm) | 145.8 | 138.5 | 145.4 | 134.4 | 124.3 | 137.3 | 131 | 125.1 | 130.7 | 136.6 |
Eucl. distance (μm) | 97.41 | 96.49 | 101.1 | 99.46 | 94.61 | 102.5 | 102.5 | 95.55 | 98.31 | 104.5 |
Branch order | 3.260 | 3.576 | 3.475 | 3.207 | 2.884 | 3.513 | 3.277 | 4.173 | 3.188 | 3.492 |
Contraction | 0.7797 | 0.8725 | 0.8214 | 0.9100 | 0.8568 | 0.858 | 0.8727 | 0.8867 | 0.8428 | 0.853 |
Partition asym. | 0.4297 | 0.4815 | 0.3786 | 0.3444 | 0.4692 | 0.4778 | 0.4314 | 0.4312 | 0.5167 | 0.3764 |
Taper | −0.0437 | −0.0438 | −0.0767 | −0.0408 | −0.0374 | −0.0925 | −0.0230 | −0.0403 | −0.0217 | −0.0048 |
Daughter ratio | 1.667 | 1.704 | 1.421 | 1.388 | 1.604 | 1.920 | 1.512 | 1.708 | 1.63 | 1.621 |
Parent daughter ratio | 0.7898 | 0.8984 | 0.8038 | 0.8959 | 0.7872 | 0.7599 | 0.8686 | 0.731 | 0.7625 | 0.6617 |
Bif. amp local (°) | 95.53 | 75.5 | 86.16 | 65.27 | 60.74 | 75.38 | 67.24 | 69.16 | 67.85 | 68.43 |
The total reconstructed dendritic surface area and volume, defined as the sum over all segments of respectively, segment surface and volume, were based on the assumption that each segment is a uniform cylinder (as opposed to a tapering one), Table
The interquartile range for overall volume is around 3000 μm3, which is about 25% of the median value. However, such variation is not unexpected, as other investigations into inter-user variance on different systems have consistently found very significant variation between both different operators on the same system and the same operator on different systems (Jaeger,
The variations in the quality of a reconstruction come from two major sources: the number of identified and segmented branches, and the quality of the segmentation of each branch (midline and radius estimation). Some gross properties, such as the total number of branches, are only attributable to one of these factors (branch identification), whereas the overall volume, for example, is a function of both.
The measurement of the dendritic radius contributing to the quality of branch segmentation is always the most variable aspect of neuronal reconstructions: due to the integration of the image volume with the Point Spread Function the edges between the dendrites and the background are blurred, and thus the choice of diameter tends to be subjective. Estimation of radii can vary significantly between different labs performing neuronal reconstruction (Scorcioni et al.,
For this experiment, participants were instructed to estimate the dendrite edge as where the brightest luminosity of the pixels first began to decrease (since the participants were working with a luminosity-inverted stack, the dendrites were lighter than the background). On the other hand, the reconstruction procedure used in (Brown et al.,
Table
In summary, these data suggest that the Neuromantic interface makes it simple to navigate the image data and identify dendrites that have not yet been segmented. This might be due to the use of an inverted image stack making dendritic details clearer or the variety of options available for overlaying the current reconstruction.
Table
Property | Neuromantic | L.Quart. | U.Quart. | NM | NL |
---|---|---|---|---|---|
Diameter (μm) | 1.726 | 1.500 | 1.950 | 1.671 | 1.519 |
Path distance (μm) | 135.5 | 127.9 | 142.0 | 125.1 | 110.2 |
Eucl. distance (μm) | 98.9 | 96.0 | 102.5 | 95.5 | 88.4 |
Branch order | 3.376 | 3.198 | 3.545 | 3.630 | 3.485 |
Contraction | 0.8574 | 0.8321 | 0.8797 | 0.8600 | 0.8704 |
Partition asym. | 0.4313 | 0.3775 | 0.4797 | 0.3949 | 0.4755 |
Taper | −0.0406 | −0.0603 | −0.0224 | −0.0257 | −0.0300 |
Daughter ratio | 1.626 | 1.467 | 1.706 | 1.641 | 1.211 |
Parent daughter ratio | 0.789 | 0.746 | 0.882 | 0.662 | 0.664 |
Bif. amp local (°) | 68.8 | 66.3 | 80.8 | 77.0 | 77.0 |
The mean diameter of the Neuron_Morpho and Neurolucida reconstructions falls within the interquartile range observed in the experiment (0.45 μm), with the proportional difference between the median and the original reconstructions being just over 10% at maximum.
The median path, and Euclidean mean, distances for branches, however, do not fall within 10% of either the Neurolucida or Neuron_Morpho reconstructions reflecting the generally larger number of branches identified by the participants.
Contraction (always between 0 and 1) is a measure of dendritic meandering, with 1 indicating perfectly straight dendrites and decreasing values increasing “wiggle.” The interquartile range of the experimental contraction values (≈0.05) encompasses the associated values of both original reconstructions. The median value observed in the experiment is also within <5% of both the original values.
For the partition asymmetry, the Neuron_Morpho and Neurolucida values again lie within the experimental interquartile range, and the median within <10% of both values. Interestingly, though, the Neuron_Morpho and Neurolucida values differ significantly by around <20% on this metric; likely attributable to the missing right hand branch in the Neuron_Morpho reconstruction (Figure
The taper measure used here is the mean decrease in diameter per unit length. The experimental reconstructions tended to have greater mean taper rates than the original reconstructions, and the interquartile range of this metric was large at 0.04.
The daughter ratio is the ratio of the radius of the wider daughter branch from a bifurcation to the thinner branch (and thus is always ≥1). Here, a large difference is observed between the Neuron_Morpho and Neurolucida reconstructions (1.64 and 1.21, respectively), with the Neuromantic interquartile range encompassing the Neuron_Morpho value but not the Neurolucida one, and the Neuromantic median being highly similar to the Neuron_Morpho value (at 1.63). It is possible that the differing interfaces bias operators into segmenting bifurcations in different ways, and Neuromantic’s basic method for semi-manual reconstruction is much more similar to Neuron_Morpho than Neurolucida’s.
As for the parent daughter ratio (the diameter of the daughter branch divided by that of the parent branch), the interquartile range includes neither the Neuron_Morpho nor the Neurolucida reconstructions. This may be partly due to a tendency for the inexperienced participants to systematically overestimate the radius leading to larger parent daughter ratio scores.
The local bifurcation angles measured in the experiment were also significantly different than the original reconstructions. The Neuron_Morpho and Neurolucida values were both highly similar (with less than a degree’s difference), and are encompassed by the experimental interquartile range (≈14°). As the bifurcation angle is only calculated based upon the angle between the parent and daughter segments, there is significant scope for subjective difference as to the parent segment placement.
Both of the original reconstructions were segmented by the same individual, thus minimizing subjective inter-user differences. Therefore, it is to be expected that the Neuron_Morpho and Neurolucida reconstructions would be more similar to each other than to Neuromantic ones.
Some of the reconstruction variability may be attributed to the fact that the participants were non-experts, hence reflecting some of the participants’ incorrect understanding of the actual task, rather than issues with the application itself (based on a short debriefing session afterward). Each participant increased in speed over the course of the experiment as they became more used to visually interpreting the image stacks and repeating the basic process of segmenting the neurons.
The results reported illustrate the effect of modifying the cost function on the
Table
Cost function | Trace 1 | Trace 2 | Trace 3 | Trace 4 | Trace 5 | Trace 6 | Trace 7 | Trace 8 | Trace 9 | Trace 10 |
---|---|---|---|---|---|---|---|---|---|---|
4.8707 | 4.5262 | 2.2036 | 5.6949 | 3.8731 | 41.6660 | 3.6509 | 164.3688 | 159.4037 | 254.2347 | |
4.7795 | 3.3969 | 2.2779 | 5.6806 | 3.6484 | 41.3370 | 3.5962 | 164.3757 | 151.8812 | 204.9152 | |
3.0021 | 4.1559 | 2.2034 | 5.3960 | 3.3608 | 35.2498 | 2.3363 | 142.2176 | 177.5774 | 199.9689 | |
3.4211 | 3.3865 | 2.1943 | 5.6674 | 3.6418 | 41.2335 | 3.5976 | 164.6490 | 141.2616 | 260.3797 | |
4.0023 | 4.0240 | 1.8946 | 4.8941 | 3.2774 | 40.1589 | 2.3352 | 143.2290 | 174.8229 | 169.0122 | |
2.5819 | 3.8798 | 2.1710 | 5.0038 | 3.5078 | 40.1783 | 3.7550 | 118.5361 | 24.8087 | 199.9920 | |
4.8383 | 3.5333 | 2.0861 | 5.6070 | 3.5499 | 41.9231 | 3.4952 | 165.1431 | 150.0178 | 265.7633 | |
2.6480 | 3.9194 | 2.2710 | 5.4042 | 3.3421 | 30.9186 | 2.4176 | 65.7683 | 25.5060 | 160.7103 | |
2.6324 | 3.5679 | 2.1196 | 5.3163 | 3.0866 | 35.6291 | 1.8305 | 39.9139 | 7.5303 | 204.9826 |
Significant variation is observed between cost functions. Particularly, much larger errors are seen in the second set of benchmarks 6–10, as it is possible for the routing algorithms to miss out significant sections of the neurite because of their meandering nature. Figure
Table
Cost | Mean rank | Overall rank | Reject |
|
---|---|---|---|---|
7.8 | 9 | 0.1128 | No | |
6.5 | 8 | – | No | |
4.7 | 5 | 0.0860 | No | |
5.8 | 6 | 0.4404 | No | |
3.8 | 3 | 0.0266 | No | |
4.0 | 4 | 0.0163 | No | |
6.2 | 7 | 0.8482 | No | |
3.6 | 2 | 0.0133 | No | |
2.5 | 1 | 0.0014 | Yes |
As expected, the neuriteness term exponent,
Table
Cost | Trace 1 | Trace 2 | Trace 3 | Trace 4 | Trace 5 | Trace 6 | Trace 7 | Trace 8 | Trace 9 | Trace 10 |
---|---|---|---|---|---|---|---|---|---|---|
−1.96 | 0.41 | −3.61 | −0.74 | −0.56 | −11.05 | −0.80 | −22.65 | −17.32 | −27.70 | |
−2.30 | 0.28 | −3.31 | −1.53 | −1.35 | −11.43 | −1.01 | −23.20 | −16.73 | −23.53 | |
−2.46 | 0.30 | −3.15 | −0.68 | −1.29 | −9.23 | −0.56 | −16.79 | −15.51 | −21.52 | |
−1.46 | 0.27 | −3.48 | −1.52 | −1.31 | −11.44 | −1.00 | −22.65 | −16.44 | −27.70 | |
−2.70 | −0.11 | −3.06 | −0.55 | −1.32 | −10.21 | −0.51 | −15.98 | −15.53 | −22.75 | |
−2.05 | −0.02 | −3.05 | −0.83 | −1.40 | −10.71 | −0.96 | −10.86 | −8.96 | −21.81 | |
−2.30 | 0.43 | −3.28 | −1.54 | −1.38 | −11.26 | −1.11 | −22.67 | −16.73 | −27.76 | |
−1.83 | −0.52 | −2.83 | −0.83 | −1.39 | −8.54 | −0.48 | −6.30 | −8.84 | −21.94 | |
−2.12 | −0.33 | −3.05 | −0.82 | −1.33 | −9.93 | 0.62 | −11.09 | −4.20 | −23.30 |
As expected from previous work, the actual length tends to be underestimated, due to Dijkstra’s algorithm preference for physically shorter routes. For the normal dendrites (benchmarks 1–5), though, the errors tend to be consistently less than 4% of the overall length. When considering the meandering dendrites, however, the length estimation errors are generally much larger, sometimes up to 20% of the overall length, which is much less acceptable and would have a very significant effect on simulation if left uncorrected. Such large errors, as explained previously in relation to
Neuromantic is a freeware application for producing three dimensional reconstruction of neurons. Its performance was demonstrated in manual and semi-automated reconstructions from non-deconvolved Transmitted Light Brightfield (TLB) image stacks. In these cases, lighting intensity varies across the image, making the data unsuitable for global thresholding to segregate dendrites from background. Also, the numerous out-of-focus artifacts mean that the data is not a true 3D voxel representation of the neuron. Significantly more image processing is thus required than for confocal stacks to extract accurate neuronal morphology.
Non-deconvolved stacks were considered, as effective deconvolution is often difficult on Golgi stained or Biocytin labeled and stained stacks. However, Neuromantic may be applied equally well to the reconstruction of dendrites from deconvolved image stacks.
The application was compared to a similar freeware system, Neuron_Morpho, and a commercially available package, Neurolucida, indicating comparable speed of use and inter-user variation consistent with that reported for other comparable studies.
Our informal survey of Neuromantic users indicates appreciation of its lightweight feel and simple interface for basic visualization and editing. The ability of dynamic image loading offers possibility to work smoothly with very large image stacks with moderate and widely available computer platforms. Semi-automated reconstruction from any given point requires simply clicking on an existing point and tracing the branch without the need, common in other systems, to open context menus to label points. Similarly, the diverse and user-friendly options of automated point selection, offer a quick way of investigating alternative reconstructions which often require modification of already completed reconstructions. For example, Neuromantic enables easy connectivity changes and immediate visual edits. Both operations require
The perceived optimal trade-off between utility and ease of learning was also a factor that motivated the selection of Neuromantic as the official editing tool in the DIADEM challenge (
To address the known problems with inter-user variance on semi-manual reconstructions, the Neuromantic 3D image stacks extension to semi-automatic tracing (Meijering et al.,
The method was evaluated in terms of reconstruction consistency, examining the effect of the routing algorithm’s cost function form on the accuracy of dendrite midlines over a range of benchmarks. Increasing the exponents of the cost function two terms significantly improved tracing quality. The term relating to the likelihood of a given pixel belonging to a dendrite was significantly more important than the term relating to directional flow.
The modification to the Dijkstra cost function, suggested by these results, produced a consistent improvement in tracing accuracy, allowing the application to automatically deal with more complex cases such as meandering dendrites. Furthermore, it also reduced required user interaction, thus decreasing the overall time needed to generate accurate 3D neuronal models.
The semi-automatic mode uses just three parameters of which only one (the standard deviation of the steerable Gaussian) requires adjustment based on the widths of reconstructed dendrites, with minimal effect on reconstruction quality, while default value of the remaining two parameters provide overall accurate reconstructions. The small number of parameters and ease of their setting is consistent with recently recommended good practice in neurite reconstruction algorithm design (Meijering,
To conclude, Neuromantic is suggested as a useful open source tool for reconstructing dendritic trees. It provides great flexibility and a good balance between speed of operation and resultant quality. Neuromantic thus is a useful addition to the repertoire of available tools for neuronal reconstruction that might appeal to some researchers.
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.
This work was supported by EPSRC Grant GR/S55897/01 and EP/F033036/1 to Slawomir J. Nasuto. Giorgio A. Ascoli acknowledges support from NIH R01 NS39600. We thank Kerry Brown for providing the
1
2