Direct Conditional Score Modeling for Accelerated MRI Reconstruction

Accelerated MRI reconstruction from undersampled k-space data is a crucial inverse problem for reducing scan times. Recent state-of-the-art methods for MRI reconstruction often employ diffusion-based generative models, but face challenges in incorporating measurement information through likelihood-based updates. This paper presents a novel approach called Direct Conditional Score (DCS) modeling that directly models the conditional score function, eliminating the need for separate likelihood terms. The method introduces a conditioning mechanism that incorporates measurement information directly into the score estimation process, aiming for more accurate and efficient reconstructions. Evaluated on the Stanford Knee MRI with Multi-Task Evaluation (SKM-TEA) dataset, the proposed approach demonstrates improved performance across various undersampling patterns and acceleration factors compared to existing methods in terms of PSNR and SSIM. The method shows generalization capability, maintaining performance on undersampling conditions different from its training data. An ablation study examines the effectiveness of the proposed conditioning approach. The proposed method’s ability to handle different undersampling settings without retraining suggests potential for more flexible MRI acquisition protocols.