PhD Defense: Segmentation Strategies for Connectomics
A human brain is estimated to have roughly 100 billion neurons connected through more than 100 thousand miles of axons and hundreds of trillions of synaptic connections. The full neural circuit within a brain is called its connectome, and understanding how it works and enables cognition, consciousness, or intelligence are important open questions in science. Recent developments in high-throughput electron microscopy imaging have enabled biologists to visually inspect brain tissue at resolutions of a few nanometers per voxel, enough to enable the analysis of neural circuits. However, the amount of data one would need to annotate to identify and reconstruct even a small circuit makes manual reconstruction efforts prohibitive.
In this thesis, we explore several computational strategies to facilitate the semi-automatic and automatic reconstruction of neurons from 3D stacks of connectomic images. We first propose Active Ribbons, a method based on deformable models and level set methods for tracing individual neurites that is amenable for interactive segmentation. We show that, unlike conventional level set methods, Active Ribbons can reliably capture neural membranes on electron microscopy stacks.
We then explore statistical models for automatic segmentation. We study the connection between the automatic segmentation of video and the reconstruction of connectomic stacks, and introduce Multiple Hypothesis Video Segmentation (MHVS), a method for the on-line segmentation of image sequences using long-term trajectories of 2D segments as possible labels. We demonstrate the applicability of MHVS in videos with an unknown number of objects and varying complexity. Building on the experience with MHVS, we propose Segmentation Fusion, a method for the automatic segmentation of connectomic stacks that does not require the explicit discovery of labels a priori and that outperforms the state-of-the-art in automatic neuron reconstruction. We finally discuss several scaling strategies for distributed neuron reconstruction and show what we think are the largest neuron reconstruction results ever obtained in connectomics.