SegMatch3D: Fast and Accurate 3D Image Segmentation by Matching Topological Features

SegMatch3D

Abstract

The proposed project aims to extend the significant improvements in topology-preserving segmentation of medical imagery from the 2-dimensional setting to volumetric image data. Following a successful Python-based proof-of-concept implementation of our method of induced matchings, a key part of the foundational algebraic stability theorem of persistent homology, we will develop a high-performance C++ implementation, meeting the efficiency demands for training CNNs on large data sets. Incorporating optimizations from the state-of-the-art persistent homology software Ripser (developed by PI Bauer), the project will make matching of persistent homology features applicable to learning-based 3d image segmentation methods for blood vessel networks as well as tracking of cell nuclei.