A semi-automated approach to dense segmentation of 3D white matter electron microscopy

Purpose Neuroscience methods working on widely different scales can complement and inform each other. At the macroscopic scale, magnetic resonance imaging methods that estimate microstructural measures have much to gain from ground truth validation and models based on accurate measurement of that microstructure. We present an approach to generate rich and accurate geometric models of white matter microstructure through dense segmentation of 3D electron microscopy (EM). Methods Volumetric data of the white matter of the genu of the corpus callosum of the adult mouse brain were acquired using serial blockface scanning electron microscopy (SBF-SEM). A segmentation pipeline was developed to separate the 3D EM data into compartments and individual cellular and subcellular constituents, making use of established tools as well as newly developed algorithms to achieve accurate segmentation of various compartments. Results The volume was segmented into six compartments comprising myelinated axons (axon, myelin sheath, nodes of Ranvier), oligodendrocytes, blood vessels, mitochondria, and unmyelinated axons. The myelinated axons had an average inner diameter of 0.56 μm and an average outer diameter of 0.87 μm. The diameter of unmyelinated axons was 0.43 μm. A mean g-ratio of 0.61 was found for myelinated axons, but the g-ratio was highly variable between as well as within axons. Conclusion The approach for segmentation of 3D EM data yielded a dense annotation of a range of white matter compartments that can be interrogated for their properties and used for in silico experiments of brain structure. We provide the resulting dense annotation as a resource to the neuroscience community.