BrainCHEF: Cross-Level Hypergraph Enhanced Fusion model for brain networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Data Science

Background:

• Modeling dynamic functional brain networks is crucial for understanding brain mechanisms.
• Graph neural networks (GNNs) show promise but struggle with data scarcity, limited supervision, and capturing high-order network structures.

Purpose of the Study:

• To address the limitations of existing methods in analyzing dynamic functional brain networks.
• To propose a novel framework, BrainCHEF, that enhances the capture of spatiotemporal characteristics and improves model generalization and interpretability.

Main Methods:

• Developed a Cross-Level Hypergraph-Enhanced Fusion Framework (BrainCHEF) integrating hypergraphs and line graphs.
• Employed hypergraph attention networks for adaptive node dependency learning and self-supervised feature masking for hyperedge interaction.
• Incorporated persistent homology analysis for fMRI signal processing and cross-level interaction mechanisms for global information integration.

Main Results:

• BrainCHEF demonstrated superior performance on ABIDE and ADHD datasets, outperforming state-of-the-art methods.
• The framework successfully identified disease-related biomarkers consistent with existing research.
• Ablation studies validated the effectiveness of hypergraph modeling and self-supervised tasks.

Conclusions:

• BrainCHEF offers a powerful new tool for analyzing dynamic brain network characteristics.
• The framework enhances model generalization and interpretability, providing valuable insights for brain disease diagnosis and research.
• The study highlights the potential of integrating hypergraph theory and self-supervised learning in neuroimaging analysis.