Deep learning-based scrubbing of fMRI data and its applications on delineating infant brain functional development trajectories

Motivation: Noise and artifacts significantly corrupt fMRI data. Goal(s): It has a significant potential in enhancing research outcomes for challenging populations like children and older subjects whose data prone to have noise, facilitating reliable fMRI studies. Approach: We present a deep learning-based Automatic fMRI Scrubbing via Graph Attention (ASGA), to perform fMRI data “scrubbing” by automatically identifying and removing contaminated volumes. To achieve this, we firstly design an easy-to-implement carpet plot-based labeling tool for human labelling, which is fed to ASGA model training. By applying ASGA to two large-cohort studies (BCP and CBCP). Results: our method effectively removed noise-contaminated volumes without human interference. Impact: Compared to other fMRI data censoring approaches, ASGA is automatic, targeting on general noise and artifacts, can better enhance fMRI analysis accuracy and research outcomes, especially useful for challenging populations such as children, older subjects, and patients.