Efficient Reconstruction and Proofreading of Neural Circuits

Abstract

This thesis presents the design and implementation of software technologies for neural circuit reconstruction in light of a need for radical improvements in efficiency and scalability. It focuses primarily on reconstructing the anatomical structures and connections of neurons in electron microscopy datasets of brains and extends to other problem domains, including object analysis in still images and videos. Its main contributions include a robust mechanism for training complex segmentation algorithms, optimization of 3D convolutions for efficient processing of volumetric data and videos, design of recurrent neural networks for accurate reconstruction of 3D neurons, and an analytics framework for the automatic discerning of potential errors in segmentation and fast proofreading neural sub-graphs. Implementing these technologies has contributed to new usable tools strategically targeted towards the efficient reconstruction of neural circuits.