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

Area of Science:

• Neuroscience
• Computational Biology
• Artificial Intelligence

Background:

• Accurate cell-type mapping is crucial for understanding brain organization.
• Manual annotation of cells in volumetric images is costly and time-consuming.
• Existing neural networks struggle with high-precision segmentation of multiple cell types in a unified framework.

Purpose of the Study:

• To introduce a novel self-supervised learning framework for automatic multi-neuronal population segmentation.
• To address the limitations of existing methods in terms of precision and cost.
• To enable the construction of comprehensive whole-brain cell-type atlases.

Main Methods:

• Developed the Voxelwise U-shaped Swin-Mamba network (VUSMamba) framework.
• Employed contrastive learning and pretext tasks for self-supervised learning on unlabeled data.
• Fine-tuned the model with minimal annotations on multi-cell-type datasets from multiplexed fluorescence in situ hybridization (multi-FISH) and VISoR microscopy.

Main Results:

• VUSMamba achieved higher segmentation accuracy compared to state-of-the-art baseline models.
• The framework demonstrated reduced computational cost.
• Enabled simultaneous high-precision segmentation of glutamatergic neurons, GABAergic neurons, and nuclei.

Conclusions:

• VUSMamba offers a unified self-supervised neural network for automated cell-type segmentation.
• The framework provides a standardized pipeline for creating and analyzing whole-brain cell-type atlases.
• This approach facilitates more efficient and accurate brain mapping.