Multi-channel spatio-temporal graph attention contrastive network for brain disease diagnosis

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Dynamic brain networks (DBNs) are vital for understanding neurological disorders, but current methods struggle with higher-order spatio-temporal patterns and integrating structural priors.
• Existing sliding window approaches for DBNs analysis often overlook complex topological relationships and multi-modal data integration.

Purpose of the Study:

• To propose a multi-channel spatio-temporal graph attention contrastive network for enhanced DBNs analysis.
• To integrate structural connectivity from diffusion tensor imaging (DTI) with functional connectivity from fMRI for a multi-modal brain network representation.
• To develop a novel network architecture capable of extracting higher-order spatio-temporal topological features for improved neurological disorder diagnosis.

Main Methods:

• Constructed dynamic functional networks from fMRI data using sliding windows.
• Embedded structural connectivity from DTI into functional connectivity graphs to create multi-modal brain networks.
• Developed a multi-channel spatial attention contrastive network with intra-window and inter-window contrastive constraints to capture topological features.
• Utilized a self-attention mechanism to integrate feature embeddings across time windows for higher-order spatio-temporal analysis.
• Employed a multi-layer perceptron (MLP) for brain network classification.

Main Results:

• The proposed method achieved superior diagnostic performance on epilepsy and ADNI datasets compared to state-of-the-art approaches.
• The network successfully extracted discriminative graph features relevant to brain diseases.
• The integration of multi-modal data (fMRI and DTI) and advanced network architecture improved the identification of neurological disorders.

Conclusions:

• The multi-channel spatio-temporal graph attention contrastive network offers a powerful new tool for analyzing DBNs.
• This approach enhances the understanding of complex brain network dynamics in neurological conditions.
• The method demonstrates significant potential for clinical applications in diagnosing brain diseases.