High-speed high-quality super-resolution diffusion model via flow-matching

Motivation: Develop a fast diffusion model-based network for high-quality super-resolution to improve image resolution in MRI. Goal(s): To propose a super-resolution network for reconstructing complex GRE data, applicable to QSM or SWI. Approach: Flow-matching, advanced variant of DDPM, was adopted for super-resolution in MRI. At each step of the flow-matching, a network output is added to the current image to predict the next image. We modified the network output to impose data-consistency in K-space. Results: The proposed network successfully reconstructed high-resolution complex GRE images and QSM maps, outperforming other networks. The processing time was shortened by 10 times (only 43.9 sec for the whole-brain). Impact: The proposed data-consistency guided super-resolution network using diffusion model via flow matching demonstrates high-quality images, outperforming existing methods while shortening the processing times more than 10 times (43.9 sec vs. 8 min 4 sec).