Fuse-Former: An interpretability analysis model for rs-fMRI based on multi-scale information fusion interaction

Resting-state functional magnetic resonance imaging (rs-fMRI), as a non-invasive neuroimaging technique, is widely used in the auxiliary diagnosis of brain diseases. However, existing deep learning-based methods are often not sensitive to the exploration of multi-scale temporal features, especially lacking effective utilization of information regarding blood oxygen level dependent (BOLD) level changes over short periods of time. Hence, we propose a brain disease recognition and analysis model for rs-fMRI based on multi-scale information fusion interaction (Fuse-Former). The design of Fuse-Former adopts a global-local model architecture. The model divides the brain into different regions of interest (ROI) using an external atlas and extracts regional BOLD response information as feature inputs. The global feature extraction module extracts features from the entire sequence through window information interaction and token fusion. The local feature extraction module proposes a KL distribution attention mechanism, which effectively selects key window time series features. It closely focuses on the subtle changes in BOLD response information during rest state. Moreover, Fuse-Former designs an interpretable module based on clustering, which unsupervisedly aggregates ROI in rs-fMRI that have similar effects on disease recognition and analyzes the correlation between ROI in each cluster. Fuse- Former model attains an accuracy of 0.738 and an AUC of 0.798 on the ADNI, and an accuracy of 0.743 and an AUC of 0.808 on the ABIDE I. When compared to advanced benchmark models, it exhibits substantial performance enhancements. Through the utilization of an interpretability module, we identify that the Dorsal Attention Network, Limbic Network, and Salience/Ventral Attention Network are particularly influential in the ADNI, whereas the Visual Network and Somatomotor Network are more significant in the ABIDE I. The experimental results demonstrate that brain network connectivity patterns exhibit significant differences across various pathologies. In terms of clustering structure, the ROIs for autism spectrum disorder (ASD) exhibit a more complex feature space distribution. Code is available at https://github.com/yjy-97/Fuse-Former.