Conceptual

Dual-Branch Spatio-Temporal Transformer for fMRI Brain-Disorder Diagnosis

A transformer architecture that classifies neurodevelopmental disorders from resting-state fMRI by jointly modelling spatial and temporal structure. Effective connectivity from Granger causality drives an eigenvector-centrality reordering of brain regions, giving the transformer a stable spatial ordering; a variable-window temporal module with cross-window attention captures multiscale dynamics of the BOLD time series; and parallel spatial and temporal branches are fused for classification. The concept covers how injecting network-derived spatial order and multiscale temporal windows into attention improves both accuracy and interpretability over connectivity-only models.