EM-Compressor: Electron Microscopy Image Compression in Connectomics with Variational Autoencoders

Abstract

The ongoing pursuit to map detailed brain structures at high resolution using electron microscopy (EM) has led to advancements in imaging that enable the generation of connectomic volumes that have reached the petabyte scale and are soon expected to reach the exascale for whole mouse brain collections. To tackle the high costs of managing these large-scale datasets, we have developed a data compression approach employing Variational Autoencoders (VAEs) to significantly reduce data storage requirements. Due to their ability to capture the complex patterns of EM images, our VAE models notably decrease data size while carefully preserving important image features pertinent to connectomics-based image analysis. Through a comprehensive study using human EM volumes (H01 dataset), we demonstrate how our approach can reduce data to as little as 1/128th of the original size without significantly compromising the ability to subsequently segment the data, outperforming standard data size reduction methods. This performance suggests that this method can greatly alleviate requirements for data management for connectomics applications, and enable more efficient data access and sharing. Additionally, we developed a cloud-based application named EM-Compressor on top of this work to enable on-thefly interactive visualization: https://em-compressor-demonstration.s3.amazonaws.com/EM-Compressor+App.mp4.