3D-DOPAMIN: Diffusion-Optimized Projection Algorithm for 3D Multi-Coil MRI Reconstruction

Abstract

Accelerating 3D MRI acquisition by leveraging multi-coil and undersampled k-space measurements offers the potential to reduce MRI scan time. However, the reconstruction process relies on robust algorithms to ensure high-fidelity images from incomplete data, with methodologies evolving from model-based methods to learning-based generative approaches, powered by the increase in data availability. The MR images generated from trained diffusion models often suffer from limited fidelity to the acquired measurements due to the lack of explicit enforcement of data consistency and physical constraints. To overcome these limitations, we propose 3D-DOPAMIN, a hybrid optimization framework for 3D multi-coil MRI reconstruction that integrates physics-based constraints with a 3D diffusion model for the inference. We demonstrate the robustness of our method across Cartesian and radial undersampling patterns with fastMRI knee datasets under various acceleration factors.