BrainCL: Transformer-Based Brain Network Contrastive Learning with Multi-Order Topology and Salience Masking

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) is vital for diagnosing neurological disorders.
• Transformers show promise in brain network analysis but struggle with topological complexity and limited region focus.
• Existing models exhibit high sensitivity to site differences and inter-subject variability, impacting performance.

Purpose of the Study:

• To enhance Transformer-based brain network analysis for improved neurological disorder prediction.
• To address limitations in capturing complex brain network topology and reliance on a few regions of interest (ROIs).
• To improve model robustness and cross-site generalization in brain network analysis.

Main Methods:

• Proposed a multi-order topology-aware Transformer (MoT-Former) incorporating preferential random walk scheme (PRWS) and hop-wise gated attention (HWGA) for multi-hop interactions.
• Introduced a salience-informed dynamic masking strategy to encourage mining subtle abnormalities from less activated ROIs.
• Developed synergistic dual-level contrastive learning to enhance representation invariance and create a class-discriminative feature space.

Main Results:

• BrainCL significantly outperformed existing methods in neurological disorder prediction.
• Achieved excellent cross-site generalization, demonstrating robustness to variations.
• Visualizations indicated BrainCL utilizes information from a broader range of ROIs.

Conclusions:

• BrainCL offers a robust and effective framework for Transformer-based brain network analysis.
• The proposed methods enhance the capture of complex topological properties and subtle functional abnormalities.
• BrainCL's ability to leverage diverse ROIs may facilitate the discovery of novel biomarkers.