Self-supervised learning analysis of multi-FISH labeled cell-type map in thick brain slices.

Introduction: Accurate mapping of the spatial distribution of diverse cell types is essential for understanding the cellular organization of brain. However, the cellular heterogeneity and the substantial cost of manual annotation of cells in volumetric images hinder existing neural networks from achieving high-precision segmentation of multiple cell-types within a unified framework.

Methods: To address this challenge, we introduce a self-supervised learning framework, Voxelwise U-shaped Swin-Mamba network (VUSMamba), for automatic segmentation of multiple neuronal populations in 300 μm thick brain slices. VUSMamba employs contrastive learning and pretext tasks for self-supervised learning on unlabeled data, followed by fine-tuning with minimal annotations. As a proof of concept, we applied the framework to a multi-cell-type dataset obtained using multiplexed fluorescence in situ hybridization (multi-FISH) combined with high-speed volumetric microscopy VISoR.

Results: Compared to state-of-the-art baseline models, VUSMamba achieves higher segmentation accuracy with reduced computational cost. The framework enables simultaneous high-precision segmentation of glutamatergic neurons, GABAergic neurons, and nuclei.

Discussion: This work presents a unified self-supervised neural network framework that offers a standardized pipeline for constructing and analyzing whole-brain cell-type atlases.