Site Common Information-Guided Site-to-Individual-to-Global Multi-View GCN for Psychiatric Diagnosis

Magnetic resonance neuroimaging (MRI)-based brain functional connections show significant potential in aiding the diagnosis and assessment of psychiatric disorders. Current neuroimaging studies utilize substantial amounts of data, frequently from multiple centers, to enhance the confidence of the results. However, multi-center neuroimaging studies face challenges due to inter-site variability caused by differences in scanner configurations and acquisition protocols, leading to low data consistency with limited generality. To address these challenges, we propose a novel site common information guided Site-to-Individual-to-Global Multi-View graph convolutional network GCN (sigGCN) framework. SigGCN extracts the shared site common information to mitigate site-effects, constructs individual multi-view representations, and integrates individual and site-common features to generate global Chebyshev networks. Results show that sigGCN achieves the highest accuracy in diagnosing autism comparing with other exiting models. Moreover, sigGCN successfully classifies autism/schizophrenia from controls across 5 brain atlases with 3 independent datasets, demonstrating its robustness and stability. By systematically leveraging the hierarchical information among sites, individuals, and global networks, sigGCN offers a reliable new way to harmonize multi-center data to improve the performance of neuroimaging based psychiatric diagnosis.Clinical Relevance-The proposed sigGCN establishes a site-to-individual-to-global framework to effectively and reliably harmonize multi-center data and to improve the performance of neuroimaging based psychiatric diagnosis.