VoxelCompress: Learned Implicit Neural Compression of 3D Connectomics Data

Abstract

Connectomics, the study of comprehensive maps of neural connections in the brain, faces a significant computational bottleneck in data acquisition and storage. Current methods for imaging brain tissue at the required nanoscale resolution via serial section electron microscopy demand an estimated 10,000 years and require 1000 exabytes to image a complete human brain. This thesis introduces VoxelCompress, a deep learning-based compression framework that drastically reduces storage requirements while preserving high-fidelity reconstructions of volumetric electron microscopy (EM) data. Through extensive experiments, we demonstrate that VoxelCompress achieves a 2048× compression ratio and 99.95% overall compression, with efficiency far surpassing traditional lossless methods (only 1.68×) and accuracy exceeding state-of-the-art learned approaches (PSNR 38+ dB). Notably, VoxelCompress is the first generalizable model that does not require retraining between volumes, making its incorporation into existing imaging pipelines highly practical. Additionally, we show that denoising the input data significantly enhances compression performance—improving the peak signal-to-noise ratio (PSNR) by 6.64 dB and enabling a reduction in model depth from six layers to five without compromising reconstruction quality.