Feasible and Transferable 3D Quality Control Model for ASL Images

Motivation: Arterial spin labeling (ASL) is vulnerable to motion and has low signal-to-noise ratio, which may cause unusable image quality. Goal(s): To propose a feasible and transferable quality control model for 3D ASL images across multiple cohorts/sites. Approach: A novel model was proposed with a long short-term memory network for flexible volume inputs and a domain-adversarial neural network for model generalization, which was validated on two different datasets and the performance was evaluated with and without domain adaptation. Results: The model achieved 24% and 31% higher accuracy in slice-wise and volume-wise evaluation on the unlabeled target dataset compared to the one without domain adaptation. Impact: The proposed model enables flexible and transferable quality control for 3D ASL images, which may offer a valuable tool for brain studies across multiple cohorts and sites.