Assessing the use of Diffusion models for motion artifact correction in brain MRI

TL;DR

This study evaluates diffusion-model-based correction of motion artifacts in 2D brain MRI by comparing an unconditional DDPM trained on motion-free data against a supervised U-Net trained with synthetic motion artifacts. Experiments on the MR-ART dataset across sagittal, coronal, and transverse planes reveal that diffusion approaches can achieve competitive accuracy in some settings but increasingly risk generating hallucinations, with performance highly dependent on data heterogeneity and input plane. The results underscore the potential of diffusion methods while highlighting diagnostic risks and the need for hallucination-mitigation strategies and possible hybrids with supervised methods, followed by rigorous clinical validation. Overall, diffusion-based motion artifact correction shows promise but requires careful use-case selection and further methodological refinements for safe clinical deployment.

Abstract

Magnetic Resonance Imaging generally requires long exposure times, while being sensitive to patient motion, resulting in artifacts in the acquired images, which may hinder their diagnostic relevance. Despite research efforts to decrease the acquisition time, and designing efficient acquisition sequences, motion artifacts are still a persistent problem, pushing toward the need for the development of automatic motion artifact correction techniques. Recently, diffusion models have been proposed as a solution for the task at hand. While diffusion models can produce high-quality reconstructions, they are also susceptible to hallucination, which poses risks in diagnostic applications. In this study, we critically evaluate the use of diffusion models for correcting motion artifacts in 2D brain MRI scans. Using a popular benchmark dataset, we compare a diffusion model-based approach with state-of-the-art methods consisting of Unets trained in a supervised fashion on motion-affected images to reconstruct ground truth motion-free images. Our findings reveal mixed results: diffusion models can produce accurate predictions or generate harmful hallucinations in this context, depending on data heterogeneity and the acquisition planes considered as input.

Figures (4)

• Figure 1: Our implementation follows a two-phase approach. In the first phase, a diffusion model (DDPM) is trained on a dataset of motion artifact-affected images. In the second phase, the trained model is used to introduce motion artifacts into clean images, generating paired datasets. These pairs enable supervised training in the subsequent step. See Algorithm \ref{['alg:motion_correction']} for more details.
• Figure 2: DDPM motion artifact correction with different time steps $n$.
• Figure 3: Comparison of the different approaches on all three views.
• Figure 4: Example of hallucination of the DDPM.