U^2MRPD: Unsupervised undersampled MRI reconstruction by prompting a large latent diffusion model

Implicit visual knowledge in a large latent diffusion model (LLDM) pre-trained on natural images is rich and hypothetically universal to natural and medical images. To test this hypothesis, we introduce a novel framework for Unsupervised Undersampled MRI Reconstruction by Prompting a pre-trained large latent Diffusion model ( U^2MRPD). Existing data-driven, supervised undersampled MRI reconstruction networks are typically of limited generalizability and adaptability toward diverse data acquisition scenarios; yet U^2MRPD supports image-specific MRI reconstruction by prompting an LLDM with an MRSampler tailored for complex-valued MRI images. With any single-source or diverse-source MRI dataset, U^2MRPD's performance is further boosted by an MRAdapter while keeping the generative image priors intact. Experiments on multiple datasets show that U^2MRPD achieves comparable or better performance than supervised and MRI diffusion methods on in-domain datasets while demonstrating the best generalizability on out-of-domain datasets. To the best of our knowledge, U^2MRPD is the first unsupervised method that demonstrates the universal prowess of a LLDM, on magnitude-only natural images in medical imaging, attaining the best adaptability for both MRI database-free and database-available scenarios and generalizability towards out-of-domain data.