Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model

TL;DR

This work tackles MRI motion artifact removal without relying on paired clean data by guiding a pre-trained diffusion model with pixel-frequency information. It introduces alternating complementary masks and a dual-domain balance to fuse low-frequency tissue textures with high-frequency and pixel details during diffusion-based reconstruction. Across multi-tissue simulations and real-clinical radiologist evaluations, PFAD achieves superior artifact removal while preserving anatomical textures, highlighting its potential for clinical deployment. By integrating k-space-aware frequency-domain processing with diffusion-based synthesis, the method advances unsupervised artifact purification in medical imaging.

Abstract

Motion artifacts present in magnetic resonance imaging (MRI) can seriously interfere with clinical diagnosis. Removing motion artifacts is a straightforward solution and has been extensively studied. However, paired data are still heavily relied on in recent works and the perturbations in k-space (frequency domain) are not well considered, which limits their applications in the clinical field. To address these issues, we propose a novel unsupervised purification method which leverages pixel-frequency information of noisy MRI images to guide a pre-trained diffusion model to recover clean MRI images. Specifically, considering that motion artifacts are mainly concentrated in high-frequency components in k-space, we utilize the low-frequency components as the guide to ensure correct tissue textures. Additionally, given that high-frequency and pixel information are helpful for recovering shape and detail textures, we design alternate complementary masks to simultaneously destroy the artifact structure and exploit useful information. Quantitative experiments are performed on datasets from different tissues and show that our method achieves superior performance on several metrics. Qualitative evaluations with radiologists also show that our method provides better clinical feedback. Our code is available at https://github.com/medcx/PFAD.

Figures (8)

• Figure 1: Schematic of our research problem and solution. During the MRI imaging process, motion artifacts are produced due to the patient's motion. Removing artifacts via a diffusion model and alternate masks in pixel-frequency domain can help radiologists make better diagnosis.
• Figure 2: The framework of PFAD. The green area is the artifact removal process in frequency domain. This process is guided by the low-frequency information and part of the high-frequency information of the motion-corrupted image, and uses the diffusion model to generate another part of clean high-frequency information. The yellow area is the pixel domain artifact removal process, which is guided by the partial image information of the forward process. After combining the both domains results, the clean image is finally generated iteratively.
• Figure 3: Alternate complementary masks. In each iteration, the complementary masks are alternately transformed to ensure artifact removal for the entire image.
• Figure 4: Comparison of results on simulated brain images: (a) Artifact images, (b) Pix2pix, (c) CycleGAN, (d) UDDN, (e) DR2, (f) GDP, (g) PFAD (ours), and (h) the ground-truth. PSNR and SSIM values of each image are shown in the corner of images. The yellow box indicates the zoomed-in visualization area, and the blue box represents the difference heatmap compared to the ground truth. The color ranges from blue to red, indicating differences from small to large, with deeper colors representing smaller differences.
• Figure 5: Comparison of results on real images: (a) Images with real motion artifacts, (b) Pix2pix, (c) CycleGAN, (d) UDDN, (e) DR2, (f) GDP, (g) PFAD (ours). The first and third rows show the artifact images in different regions of the abdominal cavity, respectively. And the second and fourth rows show the zoomed-in artifact regions for the first and third rows, respectively. Yellow arrow indicates the location of the artifact, while red arrow points to the area with severe texture deformation.