CytoCLIP: Learning Cytoarchitectural Characteristics in Developing Human Brain Using Contrastive Language Image...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Brain region function is linked to cytoarchitecture (cell arrangement and morphology).
• Manual identification of brain regions in histological sections is labor-intensive and requires expertise.
• Automated methods are needed to streamline brain region analysis.

Purpose of the Study:

• To develop an automated approach for identifying brain regions based on cytoarchitecture.
• To introduce CytoCLIP, a suite of vision-language models for learning joint visual-text representations of brain cytoarchitecture.
• To improve the efficiency and accuracy of brain region delineation in histological analysis.

Main Methods:

• Utilized pre-trained Contrastive Language-Image Pre-Training (CLIP) frameworks to create CytoCLIP.
• Developed two model variants: one for low-resolution whole-region analysis and another for high-resolution cellular-level analysis.
• Trained models on NISSL-stained fetal brain histological sections, covering 86 regions (low-res) and 379 regions (high-res).

Main Results:

• CytoCLIP achieved a weighted F1 score of 0.87 for whole-region classification.
• The high-resolution model variant achieved a weighted F1 score of 0.91 for image tile classification.
• Demonstrated superior performance compared to existing methods in region classification and cross-modal retrieval tasks.

Conclusions:

• CytoCLIP effectively automates the identification of brain regions by cytoarchitecture.
• The developed models show strong understanding and generalization capabilities for brain cytoarchitectural patterns.
• CytoCLIP offers a promising solution for accelerating neuroscience research through efficient histological analysis.