Meta-Learning Enabled Score-Based Generative Model for 1.5T-Like Image Reconstruction from 0.5T MRI

Motivation: The high-field-like image reconstruction, mainstream efforts are primarily focused on high or ultra-high fields, lacking in the reconstruction of high-field-like images from low-field ones. Goal(s): This paper presented a model for reconstructing high-field-like MR images from low-field images with unpaired data. Approach: we execute a pairing using OT-driven CycleGAN, described as "teacher learning". Subsequently, we use a diffusion model to learn the joint distribution between high-field and low-field images, guiding the reconstruction from low-field to high-field. Results: The generation experiments of T1W and T2W surpass competing experiments, and the 3-fold acceleration experiment also demonstrates the superiority of the proposed method. Impact: The proposed method represents the first attempt in the reconstruction (acceleration and generation) of images from low-field to high-field. Its potential benefits for advancing overall healthcare standards could be significant.