Let distortion-guided restoration: a physics-informed learning framework to correct prostate diffusion MRI artifacts.
Authors
Affiliations (4)
Affiliations (4)
- Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA 90048, United States.
- Department of Bioengineering, University of California, Los Angeles, Los Angeles, CA, 90095, United States.
- Department of Imaging, Cedars-Sinai Medical Center, Los Angeles, CA, 90048, USA.
- Urology, Cedars-Sinai Medical Center, Los Angeles, CA, 90048, United States.
Abstract
Susceptibility-induced distortion in single-shot echo-planar imaging (ssEPI) prostate diffusion-weighted imaging (DWI) can degrade lesion conspicuity, apparent diffusion coefficient (ADC) estimation, and confidence in prostate MRI interpretation. Conventional correction strategies typically require additional acquisitions and may be limited in severe artifact settings. To develop a physics-informed deep learning framework, distortion-guided restoration (DGR), for acquisition-free correction of ssEPI distortions in prostate DWI. Distortion-guided restoration was proposed to learn the inverse of a physically simulated ssEPI distortion process. Clinical prostate DWI and T<sub>2</sub>-weighted (T<sub>2</sub>W) images from 408 3T MR examinations originating from 2 clinical datasets, acquired on Siemens 3T MAGNETOM Vida and Biograph mMR scanners, were used for paired distorted/undistorted training data across low-b DWI (<i>b</i> = 50 s/mm<sup>2</sup>), and scanner-provided ADC maps. The restoration network integrates a convolutional neural network (CNN)-based geometric correction module with a conditional diffusion refinement module, both guided by co-registered T<sub>2</sub>W images. Model selection and benchmarking were performed on a held-out synthetic test set (<i>n</i> = 34), against conventional algorithms using correction field mapping (FSL FUGUE and TOPUP). Model performance was evaluated in a retrospective study on 34 clinical scans with severe susceptibility artifacts drawn from the same clinical datasets, using quantitative image metrics and blinded radiologist scoring. On synthetic data, the proposed DGR model achieved the highest peak signal-to-noise ratio (PSNR) and lowest normalized mean squared error (NMSE) across low-b DWI (0.089; 95% CI, 0.072-0.105) and ADC maps (0.062; 95% CI, 0.053-0.072), outperforming FSL TOPUP and FUGUE (all <i>P</i> < .001). In the 34 clinical cases affected by susceptibility artifacts, including 16 cases with histologically confirmed lesions (16/34, 47.1%), DGR substantially improved geometric fidelity (2.74 to 3.29), identified all histologically confirmed lesions (16/16, 100%), and enhanced overall image quality and diagnostic confidence (all <i>P</i> < .001). This proof-of-concept study suggests that a physics-informed hybrid CNN-diffusion framework offers a practical and acquisition-free solution for correcting severe prostate DWI distortions. It warrants further development for clinical utility and implementation.