Contrastive Graph Pooling for Explainable Classification of Brain Networks

TL;DR

This work addresses brain-network classification from resting-state fMRI by adapting graph neural networks to fMRI-specific characteristics through a contrastive dual-attention mechanism. The method constructs a contrast graph by comparing group-level summaries and then uses a differentiable pooling scheme (ContrastPool) guided by this contrast information to produce discriminative, high-level representations. Key contributions include the dual-attention CDA block, contrast-guided pooling, extensive evaluation on five datasets across three diseases, and case studies demonstrating interpretable patterns aligned with neuroscience knowledge. The approach yields superior accuracy and reduced overfitting compared with a wide range of baselines and provides clinically meaningful insights for neurodegenerative condition analysis and potential biomarker discovery.

Abstract

Functional magnetic resonance imaging (fMRI) is a commonly used technique to measure neural activation. Its application has been particularly important in identifying underlying neurodegenerative conditions such as Parkinson's, Alzheimer's, and Autism. Recent analysis of fMRI data models the brain as a graph and extracts features by graph neural networks (GNNs). However, the unique characteristics of fMRI data require a special design of GNN. Tailoring GNN to generate effective and domain-explainable features remains challenging. In this paper, we propose a contrastive dual-attention block and a differentiable graph pooling method called ContrastPool to better utilize GNN for brain networks, meeting fMRI-specific requirements. We apply our method to 5 resting-state fMRI brain network datasets of 3 diseases and demonstrate its superiority over state-of-the-art baselines. Our case study confirms that the patterns extracted by our method match the domain knowledge in neuroscience literature, and disclose direct and interesting insights. Our contributions underscore the potential of ContrastPool for advancing the understanding of brain networks and neurodegenerative conditions. The source code is available at https://github.com/AngusMonroe/ContrastPool.

Figures (11)

• Figure 1: The architecture of ContrastPool, using Autism as an example.
• Figure 2: Confusion matrices with class-wise accuracy of ContrastPool and BrainGNN on the ADNI dataset. Each confusion matrix is obtained by adding the confusion matrices on test sets of all folds.
• Figure 3: Confusion matrices with class-wise accuracy of ContrastPool and BrainGNN on the PPMI dataset.
• Figure 4: Contrast graph visualization. The connections with the top 10 ROI-wise attention weights are highlighted.
• Figure 5: Chord diagrams of contrast graphs. Only the edges with top-20 ROI-wise attention scores are shown for better visualization. (a) ROIs related to prefrontal cortex, parietal and cingulate are highlighted for Autism. (b) ROIs related to parietal and posterior are highlighted for Alzheimer's. (c) ROIs related to temporal and ventral prefrontal cortex are highlighted for Parkinson's.