Let Distortion Guide Restoration (DGR): A physics-informed learning framework for Prostate Diffusion MRI

TL;DR

This work tackles severe susceptibility distortions in prostate diffusion MRI, especially with metallic implants, by learning to invert a forward physics-based distortion model. The authors introduce Distortion-Guided Restoration (DGR), a hybrid CNN–diffusion framework that is trained on large-scale synthetic distorted/undistorted pairs generated from distortion-free references and $B_0$ templates, guided by $T_2$-weighted anatomy. DGR demonstrates superior quantitative accuracy on synthetic ground-truth data (higher PSNR and lower NMSE) and significantly improves radiologist-rated image quality and diagnostic confidence in a clinical test cohort, outperforming acquisition-dependent methods like FSL FUGUE and TOPUP. By integrating physics-based forward modeling with conditional diffusion refinement, DGR offers a practical, acquisition-free solution for robust prostate DWI restoration, including in metal-affected imaging where conventional corrections often fail.

Abstract

We present Distortion-Guided Restoration (DGR), a physics-informed hybrid CNN-diffusion framework for acquisition-free correction of severe susceptibility-induced distortions in prostate single-shot EPI diffusion-weighted imaging (DWI). DGR is trained to invert a realistic forward distortion model using large-scale paired distorted and undistorted data synthesized from distortion-free prostate DWI and co-registered T2-weighted images from 410 multi-institutional studies, together with 11 measured B0 field maps from metal-implant cases incorporated into a forward simulator to generate low-b DWI (b = 50 s per mm squared), high-b DWI (b = 1400 s per mm squared), and ADC distortions. The network couples a CNN-based geometric correction module with conditional diffusion refinement under T2-weighted anatomical guidance. On a held-out synthetic validation set (n = 34) using ground-truth simulated distortion fields, DGR achieved higher PSNR and lower NMSE than FSL TOPUP and FUGUE. In 34 real clinical studies with severe distortion, including hip prostheses and marked rectal distension, DGR improved geometric fidelity and increased radiologist-rated image quality and diagnostic confidence. Overall, learning the inverse of a physically simulated forward process provides a practical alternative to acquisition-dependent distortion-correction pipelines for prostate DWI.

Figures (4)

• Figure 1: Physics-based forward distortion simulation for ssEPI. Non-distorted prostate DWIs (DWI) and T$_2$W (T2W-TSE) images serve as anatomical references. B$_0$ field maps from hip-prosthesis patients (B$_0$ Field-OR) are denoised and low-pass filtered using a spherical-harmonics (SH) model to generate smooth field variants (B0 Field-FT). Per-pixel off-resonance values are converted into voxel-displacement maps (PDMs) according to the ssEPI forward model and applied to DWI using a splat-based EPI simulator. The synthesized distorted DWIs (DWI-SY) naturally exhibit a spectrum of characteristic artifacts---including compression, stretching, pixel pile-up, and local signal loss---providing paired distorted/undistorted data for supervised training.
• Figure 2: Overview of the proposed inverse restoration network. (a) CNN front-end: A multi-scale encoder with deformable cross-attention, cross-modal fusion, and hybrid transformer blocks extracts geometry-aware features from distorted DWI/ADC inputs under T2-weighted anatomical guidance, producing coarse distortion-corrected outputs. (b) Conditional diffusion refinement: During the forward diffusion process, clean images are progressively noised; the reverse conditional process denoises the CNN prediction toward high-fidelity results using T2+CNN features as conditioning signals. This two-stage design enables physics-guided coarse inversion and diffusion-based texture and detail restoration.
• Figure 3: Representative synthetic example comparing distortion-correction methods. (a) Corrected low-b (b = 50 s/mm$^2$) DWIs produced by each method, with the undistorted ground truth (GT) and co-registered T2W shown for reference. (b) Corresponding voxel-wise difference maps (DWI -- GT), with normalized mean squared error (NMSE) labeled. (c) Corrected ADC maps and their difference maps (bottom row). The proposed DGR method achieves substantially lower reconstruction error and improved geometric fidelity compared with the baseline, FUGUE, and TOPUP.
• Figure 4: Radiologist scores for original versus DGR-processed prostate diffusion images. Boxplots show 5-point Likert ratings (1 = poor, 5 = excellent) for (left) Geometric Fidelity, (middle) Image Quality, and (right) Diagnostic Confidence in 34 clinical ssEPI studies. For each subject, readers scored the original diffusion images and the corresponding DGR-corrected images. Black dots represent per-subject averages across all evaluated slices, and gray lines connect paired scores from the same subject to visualize individual improvements. P-values from paired t-tests demonstrate significant improvements with DGR for all three metrics.