Zero-Shot Deep Anti-Aliasing Prior for Residual Artifact Suppression in non-Cartesian k-space MRI

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging (MRI)
• Image Reconstruction

Background:

• Non-Cartesian k-space sampling in MRI can lead to residual artifacts like ringing and streaking.
• These artifacts compromise the interpretability of scanner-reconstructed images, even after preliminary corrections.

Purpose of the Study:

• To develop a zero-shot residual artifact suppression method for non-Cartesian MRI.
• To improve image quality without requiring labeled data, pre-training, or explicit degradation models.

Main Methods:

• A decoder-style generative prior was utilized, incorporating a fixed blur-kernel operator.
• The method formulates artifact suppression as an optimization problem, minimizing data-fidelity between network output and corrupted input.
• Operates directly on scanner-reconstructed images, requiring no additional learnable parameters.

Main Results:

• Demonstrated improved image quality on simulated data compared to conventional baselines.
• Achieved competitive results against supervised methods, especially under acceleration factors up to R=4.
• Reported significant improvements: up to 38% increase in SSIM and 10.64 dB increase in PSNR.
• Consistently attenuated artifacts in in vivo experiments, showing reproducible performance.

Conclusions:

• The proposed framework provides a practical, general-purpose post-processing strategy for artifact suppression in non-Cartesian MRI.
• Applicable across diverse sampling patterns and imaging settings, enhancing diagnostic accuracy.
• Offers a valuable tool for improving the reliability of MRI reconstructions.