Non-Cartesian k-space sampling in MRI is widely used, yet images reconstructed on scanners with preliminary corrections (e.g. off-resonance) often exhibit residual artifacts (e.g. ringing and streaking) that can compromise interpretation. We propose a zero-shot residual artifact suppression method that operates directly on scanner-reconstructed images without requiring labeled data, pre-training, or an explicit degradation model. The method builds on a decoder-style generative prior and incorporates a fixed blur-kernel operator that reshapes the network's inductive bias without introducing additional learnable parameters. We formulate the procedure as an optimization problem by minimizing a data-fidelity objective between the network output and the corrupted input image. We evaluate the method on simulated data and demonstrate improved image quality over conventional baselines, while remaining competitive with supervised comparisons under acceleration factors up to R = 4. Across these settings, relative to the artifact-corrupted input, SSIM improves by up to 38% and PSNR increases by up to 10.64 dB. In in vivo experiments, the proposed method consistently attenuates residual aliasing-like artifacts, indicating reproducible performance across acquisitions. Overall, the proposed framework offers a practical and general-purpose post-processing strategy for artifact suppression in non-Cartesian MRI, with applicability across diverse sampling patterns and imaging settings.
Cui et al. (Thu,) studied this question.