DeepACSON: Automated White Matter Segmentation

Abstract

Automated segmentation techniques are essential to tracing the entirety of ultrastructures in large three-dimensional electron microscopy (3D-EM) images of the brain tissue. Current automated techniques use deep convolutional neural networks (DCNNs) and rely on high-contrast cellular membranes to trace a small number of neuronal processes in very high-resolution EM datasets. We developed DeepACSON to segment large field-of-view, low-resolution 3D-EM datasets of white matter where tens of thousands of myelinated axons traverse the tissue. DeepACSON performs DCNN-based semantic segmentation and shape decomposition-based instance segmentation. With its top-down design, DeepACSON manages to account for severe membrane discontinuities inescapable with the low-resolution imaging. In particular, the instance segmentation of DeepACSON uses the tubularity of myelinated axons, decomposing an under-segmented myelinated axon into its constituent axons. We applied DeepACSON to ten serial block-face scanning electron microscopy datasets of rats after sham-operation or traumatic brain injury, segmenting hundreds of thousands of long-span myelinated axons, thousands of cell nuclei, and millions of mitochondria with excellent evaluation scores. DeepACSON quantified the morphology and spatial aspects of white matter ultrastructures, capturing nanoscopic morphological alterations five months after the injury.