Parallel and scalable neural image segmentation for connectome graph extraction

Abstract

Segmentation of images, the process of grouping together pixels of the same object, is one of the major challenges in connectome extraction. Since connectomics data consist of large quantity of digital information generated by the electron microscope, there is a necessity for a highly scalable system that performs segmentation. To date, the state-of-the-art segmentation libraries such as GALA and NeuroProof lack parallel capability to be run on multicore machines in a distributed setting in order to achieve the scalability desired. Employing many performance engineering techniques, I parallelize a pipeline that uses the existing segmentation algorithms as building blocks to perform segmentation on EM grayscale images. For an input image stack of dimensions 1024 x 1024 x 100, the parallel segmentation program achieves a speedup of 5.3 counting I/O and 9.4 not counting I/O running on an 18-core machine. The program has become I/O bound, which is a better fit to run on a distributed computing framework. In this thesis, the contribution includes coming up with parallel algorithms for constructing a regional adjacency graph from labeled pixels and agglomerating an over-segmentation to obtain the final segmentation. The agglomeration process in particular is challenging to parallelize because most graph-based segmentation libraries entail very complex dependency. This has led many people to believe that the process is inherently sequential. However, I found a way to get good speedup by sacrificing some segmentation quality. It turns out that one could trade o a negligible amount in quality for a large gain in parallelism.