MMM-DBNF: multi-atlas, multi-modal and multi-site dynamic brain network fusion for the diagnosis of schizophrenia

Dynamic brain networks (DBNs) constructed from fMRI better capture variations in functional connectivity compared to static brain functional networks and are increasingly recognized as effective biomarkers for diagnosing psychiatric disorders such as schizophrenia (SZ). However, most existing DBNs rely on a single brain atlas with single modality that ignores multi-site heterogeneity, limiting their ability to capture the brain's multi-scale organization. This study proposed a multi-atlas, multi-modal and multi-site DBN fusion method (MMM-DBNF) for SZ diagnosis. A spatial-temporal (ST) feature extraction module was developed for each atlas to integrate DBNs and brain structural networks, capturing the temporal dynamics and spatial topological features. The multi-site graph convolutional network (MGCN) modeled feature similarity and dependencies across sites, and the extracted features from all atlases were fused for classification. Experimental results showed that MMM-DBNF outperformed the other multi-atlas fusion methods on two SZ datasets (89.30% ± 0.00% and 90.16% ± 1.48%). Ablation studies highlighted the significant contributions of dynamic modeling, ST feature extraction, MGCN, and multi-atlas fusion to the improved performance. Furthermore, the identified discriminative brain regions provided insights into the functional and structural frontal-temporal dysfunction of SZ.Clinical Relevance-MMM-DBNF effectively integrates multi-atlas, multi-modal, and multi-site DBNs, providing potential clinical applications for enhancing diagnostic accuracy and identifying robust biomarkers for brain disorders.