Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model (2412.07590v2)

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.

• The paper introduces PFAD, a novel unsupervised method utilizing denoising diffusion models and complementary masks in pixel-frequency domains for motion artifact removal in MRI.
• PFAD processes images in both pixel and frequency domains simultaneously, using alternate masks and low-frequency guidance during reverse diffusion steps to restore artifact-free images.
• Experimental results across multiple datasets show PFAD outperforms existing methods (like Pix2Pix, CycleGAN) in simulated and clinical settings, demonstrating its practical applicability.

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

The paper titled "Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model" presents an innovative approach for addressing the persistent issue of motion artifacts in magnetic resonance imaging (MRI). Motion artifacts, which arise from patient movement during the scanning process, can significantly compromise the diagnostic accuracy of MRI, thus necessitating effective solutions. This paper introduces a novel unsupervised method, which they term PFAD, that leverages the power of denoising diffusion probabilistic models (DDPM), a recent advancement that has shown promise in image synthesis tasks, alongside complementary masks applied in pixel-frequency domains.

Key Methodological Contributions

1. Integration of DDPM in MRI Artifact Removal: The authors acknowledge the potential of DDPM in image generation and introduce it to the field of artifact removal. Unlike conventional GANs, the diffusion model can be guided using intrinsic data aspects from pixel and frequency domains, thus facilitating the generation of more accurate artifact-free MRI images without relying on paired clean data.
2. Pixel-Frequency Domain Approach: A significant departure from traditional artifact removal methods that operate predominantly in either the pixel or frequency domain, PFAD utilizes both domains simultaneously. The process begins with pre-training a diffusion model on unpaired clean images. By employing low-frequency components as guidance, the model reconstructs images via iterative refinements that alternately mask artifacts between pixel and frequency domains, effectively maintaining detail and texture.
3. Use of Alternate Complementary Masks: These masks play a crucial role in the proposed method, as they alternately cover different parts of image data during the reverse diffusion steps. This ensures effective disturbance of artifact structure while retaining critical image information necessary for accurate reconstruction.
4. Dynamic Parameter Adjustment: The authors introduce adjustable parameters to balance contributions from generated images and original noisy inputs, optimizing the recovery process across different image domains over time steps.

Experimental Validation and Results

Comprehensive experiments were conducted on three datasets: the Human Connectome Project (HCP), knee MRI data from fastMRI, and T2-weighted abdominal images. Across various metrics, including PSNR, SSIM, and Lpips, PFAD consistently outperformed existing methods such as Pix2Pix, CycleGAN, and other diffusion-based methods like DR2 and GDP in simulated scenarios. Moreover, radiologist evaluations underscored PFAD's superior performance in real clinical settings, indicating its practical applicability.

Implications and Future Directions

The PFAD method holds significant implications for enhancing MRI diagnostic workflows by providing a scalable, unsupervised solution to motion artifact issues. The innovative combination of pixel and frequency domain processing offers a robust framework for further exploration. Future research may explore optimizing model parameters and extending the approach to different imaging modalities or artifact conditions. As unsupervised learning continues to evolve, incorporating additional domain information, such as anatomical priors or patient-specific data, might further refine artifact correction strategies, thereby improving the coherence of recovered images.

In summary, this paper advances the methodology of artifact removal in MRI through an adept combination of modern diffusion models and dual-domain masking, setting a potential standard in the autonomous refinement of medical imaging data. This work may not only enhance the practical reliability of MRI systems but also pave avenues for improved diagnostic accuracy in clinical environments.