Abstract
Deep learning-based MRI reconstruction models have achieved superiorperformance these days. Most recently, diffusion models have shown remarkableperformance in image generation, in-painting, super-resolution, image editingand more. As a generalized diffusion model, cold diffusion further broadens thescope and considers models built around arbitrary image transformations such asblurring, down-sampling, etc. In this paper, we propose a k-space colddiffusion model that performs image degradation and restoration in k-spacewithout the need for Gaussian noise. We provide comparisons with multiple deeplearning-based MRI reconstruction models and perform tests on a well-knownlarge open-source MRI dataset. Our results show that this novel way ofperforming degradation can generate high-quality reconstruction images foraccelerated MRI.