Let distortion-guided restoration: a physics-informed learning framework to correct prostate diffusion MRI artifacts

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Susceptibility-induced distortions in single-shot echo-planar imaging (ssEPI) for prostate Diffusion-Weighted Imaging (DWI) can negatively impact lesion visibility and accuracy.
• Current correction methods often require additional MRI acquisitions and may be insufficient for severe artifacts.

Purpose of the Study:

• To develop a novel, acquisition-free deep learning framework called Distortion-Guided Restoration (DGR) for correcting ssEPI distortions in prostate DWI.
• To address limitations of conventional distortion correction techniques.

Main Methods:

• The DGR framework utilizes a physics-informed approach, learning the inverse of the ssEPI distortion process.
• It integrates a CNN-based geometric correction module and a conditional diffusion refinement module, guided by co-registered T2-weighted images.
• Training data comprised 408 clinical prostate DWI and T2W scans from Siemens scanners; performance was benchmarked against FSL TOPUP and FUGUE on synthetic and clinical data.

Main Results:

• DGR significantly outperformed conventional methods (FSL TOPUP, FUGUE) on synthetic data, achieving higher PSNR and lower NMSE for DWI and ADC maps.
• In clinical cases with severe susceptibility artifacts, DGR improved geometric fidelity, enabled 100% detection of histologically confirmed lesions, and enhanced diagnostic confidence.
• Quantitative metrics and blinded radiologist scoring confirmed the substantial improvements offered by DGR.

Conclusions:

• The physics-informed DGR framework presents a practical, acquisition-free solution for correcting severe distortions in prostate DWI.
• This proof-of-concept study highlights the potential of DGR for improving prostate MRI interpretation and warrants further clinical development.