Generative AI for Rapid Diffusion MRI with Improved Image Quality, Reliability and Generalizability

Motivation: Long scan times limit the clinical usage of diffusion MRI (dMRI)Goal(s): We aim to perform rapid dMRI with high accuracy and reproducibilityApproach: We employ a Swin UNEt Transformers (Swin) model, trained on Human Connectome Project data and conditioned on registered T1 scans, to perform generalized dMRI denoising and super-resolution, requiring only 90 seconds of scan time. Results: Compared with state-of-the-art self-supervised methods, the fully-supervised Swin UNETR achieved higher accuracy on external out-of-domain (OOD) datasets and exhibited 50% lower coefficient-of-variation for intracellular volume fraction and free water fraction measurements. Fine-tuning on even a single example scan improved performance. Impact: Our approach achieves unprecedented accuracy and reproducibility in dMRI datasets acquired in different patient populations using different scanner models and pulse sequences and will enable much shorter dMRI scan times for patients unable to cooperate with lengthy imaging protocols.