Multi-layered maps of neuropil with segmentation-guided contrastive learning

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Mapping neural circuits requires detailed cell identification, including type, subcellular components, and connectivity.
• Nanometer-resolution imaging generates vast data, but inferring cellular annotations remains a significant challenge.
• Existing methods often require extensive labeled data for accurate cell and subcellular structure analysis.

Purpose of the Study:

• To introduce a novel self-supervised machine learning technique, SegCLR, for automated analysis of neural circuit data.
• To enable accurate classification of cellular subcompartments and inference of cell types from 3D brain imagery.
• To reduce the reliance on large, manually annotated datasets for neural circuit mapping.

Main Methods:

• Developed Segmentation-Guided Contrastive Learning of Representations (SegCLR), a self-supervised learning approach.
• Applied SegCLR to 3D image volumes of human and mouse cortex.
• Utilized 3D imagery and segmentations to generate cell representations.

Main Results:

• SegCLR achieved accurate classification of cellular subcompartments.
• Performance was equivalent to supervised methods but required 400-fold less labeled data.
• Enabled cell type inference from small neural fragments (10 μm) and facilitated analysis of synaptic partners.

Conclusions:

• SegCLR offers an efficient and accurate method for annotating neural circuits from high-resolution imaging data.
• The technique enhances the utility of large-scale neural datasets, particularly those with incomplete neuronal structures.
• SegCLR supports automated, large-scale analysis of neural connectivity and cell subtypes.