Hidden Markov models for tracking neuronal structure contours in electron micrograph stacks

This paper is focused on the problem of tracking cell contours across an electron micrograph stack, so as to discern the 3D neuronal structures, with particular application to analysis of retinal images. While the problem bears similarity to traditional object tracking in video sequences, it poses additional significant challenges due to the coarse z-axis resolution which causes large contour deformations across frames, and involves major topological changes including contour splits and merges. The method proposed herein applies a deformable trellis, on which a hidden Markov model is defined, to track contour deformation. The first phase produces an estimated new contour and computes its probability given the model. The second phase detects low-confidence contour segments and tests the hypothesis that a topological change has occurred, by introducing corresponding hypothetical arcs and re-optimizing the contour. The most probable solution, including the topological hypothesis, is identified. Experimental results show, both quantitatively and qualitatively, that the proposed approach can effectively and efficiently track cell contours while accounting for splitting, merging, large contour displacements and deformations.