Automatic 3D neuron tracing using all-path pruning

Abstract

Motivation: Digital reconstruction, or tracing, of 3D neuron structures is critical toward reverse engineering the wiring and functions of a brain. However, despite a number of existing studies, this task is still challenging, especially when a 3D microscopic image has low signal-to-noise ratio (SNR) and fragmented neuron segments. Published work can handle these hard situations only by introducing global prior information, such as where a neurite segment starts and terminates. However, manual incorporation of such global information can be very time consuming. Thus, a completely automatic approach for these hard situations is highly desirable.

Results: We have developed an automatic graph algorithm, called the all-path pruning (APP), to trace the 3D structure of a neuron. To avoid potential mis-tracing of some parts of a neuron, an APP first produces an initial over-reconstruction, by tracing the optimal geodesic shortest path from the seed location to every possible destination voxel/pixel location in the image. Since the initial reconstruction contains all the possible paths and thus could contain redundant structural components (SC), we simplify the entire reconstruction without compromising its connectedness by pruning the redundant structural elements, using a new maximal-covering minimal-redundant (MCMR) subgraph algorithm. We show that MCMR has a linear computational complexity and will converge. We examined the performance of our method using challenging 3D neuronal image datasets of model organisms (e.g. fruit fly).

Availability: The software is available upon request. We plan to eventually release the software as a plugin of the V3D-Neuron package at http://penglab.janelia.org/proj/v3d.

Contact: pengh@janelia.hhmi.org.

Fig. 1.

Neuron tracing/reconstruction from images and SC examples of a reconstruction. (a) A 3D reconstruction of a fruit fly neuron. (b) The entire morphology model can typically be decomposed as individual segments (shown in different colors), which are connected at the branching points. Typically, each segment can be traced/reconstructed separately. (c) The zoom-in view of the BOX1 in (a) and (b). (d) Illustration of the modeling of image voxel information using a series of spherical SCs. The edge of image region (bright voxels) best matches to the aggregation of SCs. (e) Other types of SCs besides spheres, such as ellipsoids and cylinders, can be used in locally matching the image content and thus growing the reconstruction. (f) A maximum intensity project of a 3D stack of a fruit fly lamina neuron (courtesy of G. Rubin lab, Janelia Farm, HHMI). The neurites are highly punctuated, have high contrast in image intensity and appear to be broken. (g) Stained CA3 pyramidal neuron of a mouse brain region (courtesy of R. Tsien lab, Stanford University), where axonal varicosities make it hard to grow a reconstruction using local searching based on SCs.

Fig. 2.

Different covering situations of reconstruction nodes. Red and purple dots: leaf and non-leaf nodes, respectively. Green and blue circles: covering ranges for leaf and non-leaf nodes. (a) Keep; (b) prune; and (c) prune.

Fig. 3.

APP result. (a) A 3D image stack of a fruit fly neuron. MIP: maximum intensity projection. Arrows A and B: two locations where the intensity of the axon is very low, thus the local search using SCs would be difficult. Arrow C: a location where the bright bouton is close to another neurite tract, so that simple local SC searching can easily produce a wrong morphology. (b) ICR (red) of this neuron, where the neuron region is fully covered by reconstruction nodes. The topological connection relationship and order of these nodes are also determined in this reconstruction. (c) Pruning of the reconstruction nodes (thus also the respective SCs and redundant branches) yield a compact by representative reconstruction (red), which is still complete (the radii of the nodes are not shown for clearer illustration).

Fig. 4.

Detailed MCMR pruning process of the zoom-in area of Figure 3. The 3D display is tilted from the frontal projection in Figure 3b so that both the 3D structure of the reconstruction and its spurs are visible. (a) The remaining complete reconstruction (magenta) after pruning all dark leaf nodes repeatedly. (b) The remaining reconstruction after iterative CLP. (c) The reconstruction after an INP. (d) The mesh of the complete space-filling model of the final reconstruction in (c).

Fig. 5.

Number of remaining reconstruction nodes after each step of the MCMR pruning. Three confocal images of fruit fly neurons were used. Neuron 1 (courtesy of A. Chiang lab): 1024×1024×56 voxels, voxel XYZ size 0.215 μm ×0.215 μm ×1 μm. Neuron 2 (courtesy of G. Jefferis lab): 512×512×60 voxels, voxel size ratio Z/XY=3.03. Neuron 3 (courtesy of G. Rubin lab): 963×305×29 voxels, voxel size ratio Z/XY=1.

Fig. 6.

Consistent reconstructions produced for the same neuron from different seed locations. The final skeletons of reconstructions are intentionally displaced for better visualization. Different colors indicate different reconstructions. Dots: reconstruction nodes. Root/seed nodes are bigger dots. The major difference of these reconstructions happens around the locations of their seed locations.

Fig. 7.

Reconstructions produced for the same original neuron image, but contaminated by different levels of noise. (a) A part of the original image, where there are some dark regions that are hard for tracing. (b) A noise image where bright voxels are randomly deleted. As a result, the bouton region becomes fragmented. The noise introduced is a random deletion of q% of bright voxels (intensity >50). In this sub-figure, q=75. (c) The final skeletons of reconstructions are intentionally displaced for better visualization. Different colors indicate different reconstructions. Red: the reconstruction from noise-free image. Green, orange, yellow and magenta: q is 25, 50, 75 and 90, respectively. When q=90, most signals of the image have been removed. See summary result in Table 1 as well.

Fig. 8.

Two reconstructions produced by APP (green) and V3D-Neuron 1.0's 1-point-to-N-point automatic tracing function (red). The contours of the spherical structural components of reconstruction nodes are overlaid on top of the original image (gray scale).

Fig. 9.

An example of traced neurons that have complicated morphology. (a) The original image (courtesy of A. Chiang lab). (b) The reconstruction produced by APP.