DGST-Former: Dynamic Graph with Spatio-Temporal Transpose-Transformer for Brain fMRI Data

Time-varying brain signals can be well represented in graph-structured latent space. Dynamic Graph Neural Network integrated with Spatio-temporal Attention has been shown to have better ability of encoding meaningful representation from brain fMRI data. Although recent studies have attempted dynamic attention-based GNN model for brain fMRI data, there are still limitations in its power of expressiveness. Here, we introduce Dynamic Graph with Spatio-temporal Transpose-transformer (DGST-Former) model for encoding dynamic features from brain fMRI data. Our DGST-Former is specifically designed to embed and encode spatio-temporal representations from dynamic Functional Connectivity (FC) features. It integrates a transpose-transformer module, tailored on the basis of the transformer encoder, to capture both spatial and temporal information from dynamic graph-based data. Our proposed method has outperformed previous methods on the classification task of schizophrenia (SCZ) based on brain fMRI data. Furthermore, the analysis of encoded spatio-temporal attention are also interpretable and concord with the results of previous studies.