Unconditional Diffusion Model for 3D MRI Artifact Removal and Detail Refinement

Motivation: A considerable number of images being marred by motion artifacts. These artifacts substantially compromise the utility of MRI in clinical diagnostics and scientific research. Goal(s): To develop a robust and generalizable tool for infant brain MRI artifact reduction. Approach: We introduce an Artifact Removal (AR) method with a Detail Refinement (DR) module. The AR model employs an unconditional diffusion process trained solely on artifact-free images to improve generalization, while the DR module minimizes residual discrepancies to ensure high structural fidelity. Results: Our method effectively eliminates motion artifacts while preserving the structural integrity and fidelity of the images, surpassing the performance of popular methods. Impact: We propose a motion artifact reduction method based on unconditional DPM with a supervised fine-tuning module, DR. This approach demonstrates significant accuracy and robustness. Our method is highly valuable to neuroscience and clinical studies on existing and future large-scale datasets.