Connectome-GTC: A Unified Framework for Brain Functional and Structural Connectomes Generation, Translation, and Classification

Deep learning (DL) approaches have shown great promise in classifying brain disorders using functional connectivity (FC) and structural connectivity (SC). However, achieving optimal performance with DL-based classification typically requires large sample sizes, and collecting brain connectome data can be both difficult and expensive. Acquiring both modalities simultaneously presents additional challenges. While existing data augmentation techniques using deep generative models have demonstrated effectiveness in synthesizing new data to support DL models, most focus on a single modality and cannot infer SC from FC or vice versa. In this study, we propose a unified framework for the generation, translation, and classification of brain functional and structural connectomes (Brain-GTC). This calibrated pipeline utilizes generative models to synthesize new FC and SC. With the synthesized data, Brain-GTC employs a fusion-based classifier that integrates static FC (sFC), dynamic FC (dFC), and SC as inputs to enhance downstream classification. Notably, it can translate new FC and SC data even in the absence of one modality. Experimental results demonstrate substantial improvements, signif-icantly outperforming models without data augmentation or translation and those relying solely on sFC, dFC, or SC.