Motivation: Multi-modal MR images are vital to various disease diagnosis. Lengthy acquisition time limits its clinical applications. MRI synthesis can be an alternative to mitigating this issue. Goal(s): We aim to develop a general multi-modal MRI synthesis model capable of generating metadata-specified brain MR images from acquired scans. Approach: We compile a dataset of 31,407 3D brain MR scans. An MRI-dedicated text encoder is pre-trained to extract features from textual metadata, empowering the MR image synthesis model to precisely yield metadata-specified images. Results: Our generative foundation model provides reliable MRI sequences according to specified scanning parameters, and demonstrates superior generalizability across multi-center data. Impact: Our general multimodal MRI synthesis foundation model is capable of quickly and cost-effectively providing metadata-tailored multiple MR sequences, enabling clinicians and researchers to customize the desired MR images using this convenient AI technology, thereby enhancing diagnostic precision and efficiency.
Wang et al. (Tue,) studied this question.