Unsupervised motion artifact purification guided by joint prior from pixel and k-space domains.
Authors
Affiliations (8)
Affiliations (8)
- State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, 710071, Shaanxi, China. Electronic address: [email protected].
- Faculty of Data Science, City University of Macau, Macao Special Administrative Region, 999078, China. Electronic address: [email protected].
- Department of Radiology, Guangdong Provincial People's Hospital, Guangzhou, 510080, Guangdong, China. Electronic address: [email protected].
- Department of Radiology, Huangshi Central Hospital, Affiliated Hospital of Hubei Polytechnic University, Huangshi, 435000, Hubei, China. Electronic address: [email protected].
- Department of Radiology, Xiangyang No. 1 People's Hospital, Hubei University of Medicine, Xiangyang, 441000, Hubei, China. Electronic address: [email protected].
- Department of Radiology, Guangdong Provincial People's Hospital, Guangzhou, 510080, Guangdong, China. Electronic address: [email protected].
- State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, 710071, Shaanxi, China. Electronic address: [email protected].
- State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, 710071, Shaanxi, China. Electronic address: [email protected].
Abstract
Magnetic resonance imaging (MRI) plays an important role in clinical diagnosis, nevertheless, it suffers from motion artifacts in many real-world scenarios. Removing motion artifacts can alleviate this interference and thus facilitate precise diagnosis. However, recent works mainly rely on paired data and the artifactual perturbations in k-space (frequency domain) have not yet received sufficient attention. In this work, we propose an unsupervised purification method which incorporates k-space information of noisy MRI images as a guidance to acquire clean MRI images via a pre-trained diffusion model. Specifically, motion artifacts are mainly carried by high-frequency components in k-space, so the low-frequency components can be utilized as guidance to recover proper tissue texture. Meanwhile, considering that high-frequency and pixel information are still valuable for refining shape and texture details, we design alternate complementary masks to leverage useful information while destroying their artifactual structure. Additionally, we propose a prior meanbook as a stable denoising reference, which effectively improves the performance of artifact removal. Experiments on three datasets from different body parts are conducted to evaluate the effectiveness of the proposed method, which show that our method exhibits superior performance in multiple quantitative metrics and the qualitative clinical assessments.