A Comparison Between Boundary-Detection and Flood-Filling Approaches to EM Neuronal Segmentation

Abstract

Automatic neuronal segmentation of cross-sectional stacks of electron microscopy im- ages of the brain is a crucial step towards creating detailed 3D maps of the brain’s neuronal and synaptic structure to advance understanding of its connectivity. To solve this neuronal segmentation problem, convolutional neural networks have been used to solve a boundary-detection problem, in which the network classifies the affini- ties between pairs of adjacent pixels to being part of the same neuron or not. A connected-components algorithm such as watershed is then typically used to convert this affinity map into a segmentation. In lieu of this two-step approach, we instead build upon the flood-filling approach of Januszewski et al, which reframes the seg- mentation problem as many individual subproblems, in which the binary mask for one specific neuron is generated in its entirety by ’flooding’ from a starting pixel. The advantage of this flood-filling approach over the affinities + watershed approach is its capacity to correctly segment misaligned or thin neurons that have non-overlapping cross-sections, which affinity prediction cannot handle. We therefore explore the effectiveness of fixed-FOV flood-filling nets in correctly segmenting neurons with mis- aligned/warped slices, as well as implementing a fully recurrent, moving-FOV version of the flood-filling net for comparison against the state-of-the-art affinity-prediction approaches, and compare these approaches against traditional convolutional affinity- detection nets.