Topology-constrained graph transformer network for structural and functional brain organization

Area of Science:

• Neuroscience
• Computational Neuroscience
• Graph Theory

Background:

• The human brain's complex organization requires advanced methods for topological characterization.
• Conventional graph neural networks (GNNs) struggle with higher-order structures, while Transformers are computationally intensive.
• Existing methods often overlook the intricate topological organization of brain networks.

Purpose of the Study:

• To develop a novel Topology-Constrained Graph Transformer Network (TC-GTN) for enhanced brain network analysis.
• To integrate brain network topology explicitly into graph learning for improved efficiency and accuracy.
• To accurately characterize neuroanatomical divergence across different pathological states using brain imaging data.

Main Methods:

• Proposed TC-GTN combines cycle-constrained graph convolution for local dependencies and MST-guided Transformer for global dependencies.
• Introduced cycle-based edge positional encodings (CEPE) for topological context.
• Evaluated on large-scale structural (DWI) and functional (fMRI) brain networks from UK Biobank and ABCD datasets.

Main Results:

• TC-GTN outperformed state-of-the-art GNNs in sex classification and brain-age estimation tasks.
• The model demonstrated superior accuracy, interpretability, and generalizability across datasets.
• Clinical analysis revealed accelerated aging in multiple sclerosis and dementia, and heterogeneous alterations in stroke and Parkinson's disease.

Conclusions:

• TC-GTN offers a powerful and efficient approach for analyzing complex brain network topology.
• The model provides a valuable tool for understanding neurodevelopment, aging, and neurological disorders.
• TC-GTN's ability to characterize neuroanatomical divergence holds significant clinical potential for biomarker discovery.