An explainable deep learning framework for characterizing and interpreting human brain states

Shu Zhang, Junxin Wang, Sigang Yu, Ruoyang Wang, Junwei Han, Shijie Zhao, Tianming Liu, Jinglei Lv

Abstract: Highlights•Domain knowledge informed SAGPool is proposed to characterize the brain states.•Dense individualized connectivity-based cortical landmarks are used as graph nodes.•Dictionary learning is utilized to measure functional interactions among graph nodes.•Graph nodes which most contributed for the multitask classification can be obtained.•Learned features can elucidate the most distinguishing brain regions among tasks.