MoAR-Diff: Controlled Diffusion Model for Infant Brain MR Motion Artifact Reduction

Brain Magnetic Resonance Imaging (MRI) provides high-resolution structural details for neuroscience studies but is often affected by motion artifacts, especially in children. These artifacts substantially compromise the imaging quality and the wider utility of MRI. The efficacy of existing deep learning-based head motion artifacts eliminating approaches is often curtailed by the scarcity of paired data. To address this, we propose a novel Motion Artifact Reduction Diffusion model (MoAR-Diff). Our model allows to be pretrained on images without head motion artifacts, and fine-tuned using a smaller set of paired data. In the inference phase, we use a smartly designed denoising strategy and a frequency-domain consistency constraint to reduce artifacts while preserving the structural integrity and fidelity of the denoised images. Additionally, by incorporating age features from a pretrained text encoder, the model has the flexibility to handle rapidly developing brain for infants from 0–6 years old. This approach ensures the accuracy and robustness of the model across different age groups, surpassing the performance of both popular unsupervised and supervised methods.