Graph-level contrastive learning with self-aware and cross-sample topology augmentation for brain disorder diagnosis using rs-fMRI.

Resting-state functional MRI (rs-fMRI) is widely used for diagnosing and analyzing brain disorders. However, existing fMRI studies have shown that learning-based approaches depend heavily on labeled training data, which is difficult to obtain due to the substantial time and effort required for annotation in clinical settings. To address these challenges, we propose GCSC-TA (Graph-level Contrastive Learning with Self-aware and Cross-sample Topology Augmentation) for brain disorder diagnosis and analysis using rs-fMRI. The proposed GCSC-TA generates two complementary augmented brain networks for each subject by introducing self-aware and cross-sample topology augmentations. This dual-view strategy enhances the identification of individual-specific features and also amplifies inter-subject functional heterogeneity. Moreover, we designed a min-max contrastive loss function to accommodate augmented brain networks, overcoming the limitations of traditional projection-based methods while performing graph-level contrastive learning on the original integrity of the brain topology structure. Extensive experiments on a private Major Depressive Disorder (MDD) dataset and the publicly available Autism Spectrum Disorder (ABIDE) dataset demonstrate the superior classification performance of GCSC-TA over several state-of-the-arts. Furthermore, GCSC-TA also identifies abnormal brain connectivity patterns associated with MDD and ASD, thereby advancing the interpretability and clinical utility of rs-fMRI for clinical diagnosis.