Patch-based Conditioned Denoising Diffusion Probabilistic (PC-DDPM) for Magnetic Resonance Imaging Reconstruction

Motivation: Although MRI is a powerful medical imaging technique, its utility is limited by its prolonged scan time. Since deep learning methods for reconstructing undersampled MRI haven't achieved rapid and reliable high-resolution results, we investigated diffusion models as a potential solution. Goal(s): Improve MRI reconstruction while accelerating the inference process with diffusion model. Approach: Proposed Patch-based Conditioned Denoising Diffusion Probabilistic Model (PC-DDPM) that predicts the distribution of clean images from noisy input by conditioning on noisy patches. Results: Experiment result shows PC-DDPM outperforms U-Net, Vision Transformer, and conditioned DDPM by better reconstruction performance and shorter inference time. Impact: The implementation of patch-based conditioned DDPM for MRI reconstruction can speed up reconstruction and image acquisition while preserving the image quality. Patients could benefit from shorter scan time, and medical facilities could increase the patient throughput.