3D consistent data restoration for corrupted MRI slices using DDPMs with posterior sampling

Motivation: Cardiac CINE and EPI MRI sequences rely on minimal motion during acquisition. Irregular heartbeats and patient movement disrupt k-space sampling, resulting in signal dropout that compromises clinical assessment. Goal(s): Develop a method to restore corrupted slices in 3D MRI volumes while maintaining anatomical consistency. Approach: Implemented a 3D self-supervised network that restores corrupted slices using Denoising Diffusion Probabilistic Model with posterior sampling, which does not require corruption model during training. Results: The method achieved superior accuracy compared to conventional techniques, demonstrating lower NRMSE (0.013±0.003 vs 0.036±0.017) and higher PSNR (38.6±1.7 vs 29.8±2.7) across cardiac structures compared to Deep Image Prior and cubic interpolation. Impact: The demonstrated combination of diffusion posterior sampling with self-supervised learning establishes a framework for artifact-robust medical image restoration. This advances both computational efficiency in MRI post-processing and enables new research into automated quality assessment of cardiac functional metrics.