BrainGraphDiff: A framework for enhanced brain network analysis via adaptive subgraph generation

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Mental disorder diagnosis is a global challenge.
• Graph Neural Networks (GNNs) advance brain network analysis but face data limitations.
• Heterogeneous and scarce medical imaging data hinder model generalization and clinical application.

Purpose of the Study:

• To enhance the diversity of training data using generative models.
• To address prediction bias and poor generalization caused by data scarcity.
• To optimize brain network analysis model prediction results efficiently.

Main Methods:

• Proposed the BrainGraphDiff framework with a partial graph generation module.
• Introduced the GL-PGIB strategy for adaptive subgraph extraction.
• Utilized key graph structures as anchors for label-related subgraph scope adjustment.

Main Results:

• Demonstrated customized subgraph extraction for diverse samples.
• Showcased effective balancing of model efficiency and performance through the generative module.
• Validated the framework's ability to improve brain network analysis predictions.

Conclusions:

• The BrainGraphDiff framework and GL-PGIB strategy effectively address data heterogeneity and scarcity in brain network analysis.
• This approach enhances the reliability and scalability of diagnostic models for mental disorders.
• Customized subgraph extraction and generative modeling offer a promising solution for clinical applications.