Learnable Brain Connectivity Structures for Identifying Neurological Disorders.

Area of Science:

• Neuroscience
• Machine Learning
• Graph Neural Networks

Background:

• Brain networks are crucial for identifying neurological disorders.
• Current graph neural network models use non-learnable, predefined statistical metrics for brain network construction.
• This limits model flexibility and robustness across different brain disorders.

Purpose of the Study:

• To propose a novel, flexible module called Brain Structure Inference (BSI).
• To enable end-to-end training within a unified framework for downstream tasks.
• To learn optimal underlying graph structures directly for specific neurological disorder classification tasks.

Main Methods:

• Developed the Brain Structure Inference (BSI) module.
• Integrated BSI into a unified framework for seamless incorporation with downstream tasks.
• Enabled end-to-end training for learning task-specific brain graph structures.

Main Results:

• Achieved classification accuracies of 74.83% and 79.18% on two public datasets.
• Demonstrated at least a 3% improvement over existing state-of-the-art methods.
• Showcased high interpretability and consistency with previous findings.

Conclusions:

• BSI offers a flexible and effective approach to learning brain network structures.
• The method enhances the accuracy and robustness of neurological disorder identification.
• BSI represents a significant advancement in applying graph neural networks to neuroimaging analysis.