Predicting the retinotopic organization of human visual cortex from anatomy using geometric deep learning

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Form and function are intrinsically linked in biological systems, including the human brain.
• Cortical shape predicts retinotopic organization in the visual cortex, but current models fail to capture individual variations.
• Existing methods rely on templates, limiting their ability to model idiosyncratic structure-function relationships.

Purpose of the Study:

• To develop a novel geometric deep learning model to predict functional organization in the human visual cortex.
• To leverage individual cortical structure for more accurate and personalized predictions of brain function.
• To explore the relationship between brain anatomy and function in a non-Euclidean space.

Main Methods:

• Developed a geometric deep learning framework that exploits cortical geometry.
• Trained the model on anatomical data to predict functional organization.
• Applied the model to the human visual cortex to learn structure-function relationships.

Main Results:

• The model successfully predicted functional organization across the visual cortical hierarchy.
• The neural network captured nuanced, individual variations in structure-function relationships.
• Demonstrated the model's ability to generate realistic and idiosyncratic functional maps.

Conclusions:

• Geometric deep learning offers a powerful approach to model the complex link between brain structure and function.
• This method surpasses template-based approaches by accounting for individual anatomical differences.
• The flexible framework is applicable to diverse problems involving non-Euclidean data structures.