Contrastive Multi-View Composite Graph Convolutional Networks Based on Contribution Learning for Autism Spectrum...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) shows brain activity and is promising for autism spectrum disorder (ASD) classification.
• Existing graph convolutional network (GCN) methods for ASD classification using rs-fMRI primarily rely on low-level functional connectivities (FCs), neglecting high-level discriminative knowledge and phenotypic information, and often suffer from overfitting due to limited samples.

Purpose of the Study:

• To propose a novel contrastive multi-view composite GCN (CMV-CGCN) for enhanced ASD classification.
• To integrate both FCs and higher-order functional connectivities (HOFCs) with phenotypic information.
• To address the overfitting issue in ASD classification using GCNs.

Main Methods:

• Constructing a pair of graphs based on FC and HOFC features, incorporating shared phenotypic information.
• Developing a novel contrastive multi-view learning method focusing on consistent representations across views.
• Implementing a contribution learning mechanism to allow varying contributions from FC and HOFC features.

Main Results:

• The CMV-CGCN method achieved an accuracy of 75.20% and an AUC of 0.7338 on the ABIDE dataset (613 subjects).
• The proposed method demonstrated superior performance compared to existing state-of-the-art methods.
• The study successfully integrated multi-view data (FC, HOFC, and phenotypic information) for improved classification.

Conclusions:

• The CMV-CGCN offers a promising approach for ASD classification using rs-fMRI data.
• Integrating multi-view information and employing contrastive learning effectively enhances classification performance.
• The method shows potential for clinical application in diagnosing ASD.