Multi-Layered Maps of Neuropil with Segmentation-Guided Contrastive Learning

Abstract

Maps of the nervous system that identify individual cells along with their type, subcellular components, and connectivity have the potential to reveal fundamental organizational principles of neural circuits. Volumetric nanometer-resolution imaging of brain tissue provides the raw data needed to build such maps, but inferring all the relevant cellular and subcellular annotation layers is challenging. Here, we present Segmentation-Guided Contrastive Learning of Representations ("SegCLR"), a self-supervised machine learning technique that produces highly informative representations of cells directly from 3d electron microscope imagery and segmentations. When applied to volumes of human and mouse cerebral cortex, SegCLR enabled the classification of cellular subcompartments (axon, dendrite, soma, astrocytic process) with 4,000-fold less labeled data compared to fully supervised approaches. Surprisingly, SegCLR also enabled inference of cell types (neurons, glia, and subtypes of each) from fragments with lengths as small as 10 micrometers, a task that can be difficult for humans to perform and whose feasibility greatly enhances the utility of imaging portions of brains in which many neuron fragments terminate at a volume boundary. These predictions were further augmented via Gaussian process uncertainty estimation to enable analyses restricted to high confidence subsets of the data. Finally, SegCLR enabled detailed exploration of layer-5 pyramidal cell subtypes and automated large-scale statistical analysis of upstream and downstream synaptic partners in mouse visual cortex.