3-d Centerline Extraction of Axons in Confocal Microscopic Stacks

Extraction of geometrical features of biological structures is an active research topic. Accurate tracking algorithms provide valuable quantitative data which not only helps reduce manual labor, but also helps biologists answer a range of basic-science questions. This thesis presents a hybrid algorithm for the centerline extraction of axons in a stack of cross-sectional images acquired from a laser scanning confocal microscope. In our work, recovery of neuronal structures from such datasets helps biologists address questions regarding the pattern of synapse elimination at neuromuscular junctions in a developing muscle in mammals. Although many algorithms for centerline extraction exist in practice, none are designed for this particular application. The data acquired using fluorescence microscopy contains many artifacts such as blurred boundaries, non-uniform intensities of fluorescent radiations and the presence of noise, which make the tracking process difficult. A robust segmentation algorithm based on probabilistic region growing is introduced, which uses the shape and intensity information of the axons in the crosssections to minimize the errors in tracking. The final result of the tracking algorithm is a three dimensional centerline model. We demonstrate our algorithm on three datasets and compare its performance with the repulsive snake algorithm.