Multi-View Comparative Self-Supervised Graph Learning for Neuro-disease Diagnosis Using Functional Brain Networks

The burgeoning interest in diagnosing neurological diseases through the analysis of functional brain networks (FBNs) has been a focal point of recent research. Despite the promise of this approach, conventional graph learning techniques are often constrained by their reliance on extensive datasets annotated with clinical information. To transcend this constraint, we introduce a novel self-supervised graph learning model, termed Multi-View Comparative Self-Supervised Graph Learning (MCS-GL), which leverages Personalized PageRank (PPR) to increase samples and uses local-global information as multi view for comparative learning to enhance FBN-based diagnostics. With this approach, a pre-trained model is designed to generate information-rich Functional Magnetic Resonance Imaging(fMRI) representations from unlabeled training data. Fine-tuning the model for downstream disease identification tasks significantly reduces the requirement for large amounts of labeled training data. We have rigorously evaluated the efficacy of our MCS-GL model through its application to an autism diagnosis task. The experimental outcomes not only substantiate the superiority of our model but also underscore the pivotal role of self-supervised learning in advancing FBN-based diagnostics of neurological diseases.