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

Jiayu Ye, Yanting Li, An Zeng, Dan Pan

Abstract: Highlights•Fuse-Former constructs a block-wise feature extraction framework from global to local, directly processing the BOLD response in fMRI through multi-scale information fusion, achieving targeted high-order BOLD response interaction.•Fuse-Former proposes a KL distribution attention mechanism to obtain key temporal window features in short time frames, exploring feature segments with more significant BOLD level changes over shorter periods, thereby improving the accuracy of brain disease recognition.•Fuse-Former designs an interpretable module based on clustering to allocate ROI with similar functions, explore the correlation between brain regions, and obtain correlation scores of ROI within clusters.•Fuse-Former demonstrates its outstanding performance and reliable interpretability in different tasks of ADNI and ABIDE I. The results prove that Fuse-Former is highly competitive in the field of rs-fMRI analysis.