MoCo-Diff: Adaptive Conditional Prior on Diffusion Network for MRI Motion Correction

Feng Li, Zijian Zhou, Yu Fang, Jiangdong Cai, Qian Wang

Abstract: Magnetic Resonance Image (MRI) is a powerful medical imaging modality with non-ionizing radiation. However, due to its long scanning time, patient movement is prone to occur during acquisition. Severe motions can significantly degrade the image quality and make the images non-diagnostic. This paper introduces MoCo-Diff, a novel two-stage deep learning framework designed to correct the motion artifacts in 3D MRI volumes. In the first stage, we exploit a novel attention mechanism using shift window-based transformers in both the in-slice and through-slice directions to effectively remove the motion artifacts. In the second stage, the initially-corrected image serves as the prior for realistic MR image restoration. This stage incorporates the pre-trained Stable Diffusion to leverage its robust generative capability and the ControlUNet to fine-tune the diffusion model with the assistance of the prior. Moreover, we introduce an uncertainty predictor to assess the reliability of the motion-corrected images, which not only visually hints the motion correction errors but also enhances motion correction quality by trimming the prior with dynamic weights. Our experiments illustrate MoCo-Diff’s superiority over state-of-the-art approaches in removing motion artifacts and retaining anatomical details across different levels of motion severity. The code is available at https://github.com/fengza/MoCo-Diff.