CODE: A SELF-SUPERVISED CONSISTENCY MODEL FRAMEWORK FOR MRI DENOISING

Magnetic Resonance Imaging (MRI) is great for visualizing soft tissues and quantitative assessment of tissue properties, but its acquisitions are often limited by signal noise originating from the patient's thermal and physiological processes, motion, and hardware limitations/imperfections. We propose CoDe, a self-supervised consistency model (CM) framework that achieves efficient one-step MRI denoising. Our approach consists of two components: a noise estimation model that predicts the noise level of the input image and a CM performing one-step denoising to recover clean images. Additionally, we introduce a Random Matrix Theory (RMT)-based regularization that leverages physical noise statistics to enhance structural fidelity. Experiments on public brain and in-house prostate diffusion MRI datasets demonstrate that CoDe achieves superior image quality compared with existing methods while maintaining fast, one-step inference. Code is available at: https://github.com/JimmyHou123/CoDe-model.