Contrastive Graph Pooling for Explainable Classification of Brain Networks.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for measuring neural activation and identifying neurodegenerative diseases like Parkinson's, Alzheimer's, and Autism.
• Analyzing fMRI data using graph neural networks (GNNs) requires specialized designs due to the unique characteristics of fMRI data.
• Developing GNNs that generate effective and domain-explainable features for brain networks remains a significant challenge.

Purpose of the Study:

• To propose a novel GNN approach, ContrastPool, tailored for fMRI data analysis.
• To enhance the utilization of GNNs for brain networks by addressing fMRI-specific requirements.
• To improve the extraction of effective and explainable features from fMRI data for neurodegenerative condition research.

Main Methods:

• Introduced a contrastive dual-attention block for GNNs.
• Developed a differentiable graph pooling method named ContrastPool.
• Applied the proposed method to 5 resting-state fMRI brain network datasets across 3 diseases.

Main Results:

• Demonstrated the superiority of ContrastPool over state-of-the-art baseline methods.
• Validated the extracted patterns against existing neuroscience domain knowledge.
• Confirmed that ContrastPool provides direct and insightful findings for neurodegenerative conditions.

Conclusions:

• ContrastPool offers a powerful tool for advancing the understanding of brain networks.
• The method has significant potential for improving the analysis of neuroimaging data in neurodegenerative disease research.
• The proposed approach facilitates the discovery of novel insights into brain function and dysfunction.