Optimizing the Computational Efficiency of 3D Segmentation Models for Connectomics

EasyChair Preprint 2500

Abstract

The field of connectomics aims to map the interconnections between biological neurons within nervous systems at the scale of single synapses to gain insights into the structure and functional organization of biological neural networks. A critical task for the success of the connectomics enterprise is the segmentation of neurites from high precision electron microscopy (EM) images. This task requires models to be both accurate and computationally efficient in order to process large volumes of very high precision microscopy images. In recent years, deep learning based models have become very accurate at this task, at the cost of being very computationally intensive. In this paper, we analyse the computational efficiency of one such successful model and identify several computational bottlenecks. We propose different optimizations to increase the computational efficiency of this model and achieve a 5 times speed up in computation time while slightly improving on the baseline model accuracy.

Keyphrases: 3D CNN, Neurite segmentation, computational efficiency

@booklet{EasyChair:2500,
  author    = {Weihao Zhuang and Hascoet Tristan and Ryoichi Takashima and Tetsuya Takiguchi and Yasuo Ariki},
  title     = {Optimizing the Computational Efficiency of 3D Segmentation Models for Connectomics},
  howpublished = {EasyChair Preprint 2500},
  year      = {EasyChair, 2020}}