AI Model Improves Prostate MRI Distortion Correction

Geometric distortion in prostate diffusion-weighted MRI can make anatomy appear shifted or deformed, reducing alignment with T2-weighted images and complicating interpretation. A 2026 analysis in European Radiology Experimental evaluated DeDistortNet, an AI model designed to correct these distortions without extra MRI sequences or paired distorted and undistorted scans. The model combines information from distorted diffusion-weighted images with anatomical guidance from T2-weighted images. Testing with simulated and clinical distortion showed the clearest gains in severe cases, especially in the peripheral zone.

Distortion Remains a Challenge in Prostate MRI

Prostate diffusion-weighted imaging is often acquired using single-shot echo-planar imaging. This technique can be affected by susceptibility artefacts, which can shift or deform the prostate image. As a result, diffusion-weighted images may not match the anatomy shown on T2-weighted images. This mismatch can reduce confidence in image interpretation and can also affect image analysis.

Several existing methods can reduce distortion. Some use extra imaging to calculate how much anatomy has shifted. Others acquire images in opposite directions to estimate the distortion pattern. Different acquisition techniques can also reduce distortion during scanning. These methods can improve image quality, but they may require extra scan time or specialised protocols. Patient motion can also make longer or more complex imaging harder to use in routine prostate MRI.

Image registration offers another option. It attempts to align distorted diffusion-weighted images with T2-weighted images. This approach does not require extra acquisitions, but severe distortion can still cause problems. Signal pile-ups and signal voids can make alignment unreliable. DeDistortNet takes a different approach by generating corrected diffusion-weighted images. It uses routine prostate MRI inputs and aims to restore anatomical consistency without changing the imaging protocol.

The Model Uses Image Signal and Anatomy

DeDistortNet combines two types of guidance. The distorted diffusion-weighted image provides information about the image signal. The T2-weighted image provides anatomical structure. This combination helps the model keep the diffusion-related image information while following the prostate anatomy seen on T2-weighted MRI. The aim is to produce corrected diffusion-weighted images that are closer to the expected anatomical shape.

Must Read: MRI Radiomics Refines Prostate Surveillance Choices

Training used the public PROSTATEx dataset, which included prostate MRI exams from 328 male subjects. Images without visible distortion were used for training and validation. Exams with distorted slices were used for testing. Routine clinical prostate MRI does not provide matched distorted and undistorted diffusion-weighted images, so training used simulated distortion. Undistorted images were deliberately altered to create realistic examples of image warping and signal disruption.

The simulation process included local deformation near the posterior prostate, where rectal-adjacent artefacts can affect image quality. It also included signal changes resembling pile-ups or voids, as well as small shifts and rotations. This allowed the model to learn from varied distortion patterns. Corrected images were then used to generate outputs used in prostate MRI interpretation, including apparent diffusion coefficient maps. The method therefore remains linked to standard prostate MRI review rather than requiring a separate imaging workflow.

The Greatest Gains Appear in Severe Distortion

Testing used both simulated distortion and clinically distorted images. In simulated data, DeDistortNet improved image quality across different distortion levels and performed better than comparison generative models. The improvements were strongest in the peripheral zone, where severe and extreme distortion showed better recovery of image quality and structure. The model also reproduced distortion-free images when no artefact was present.

Clinically distorted images could not be compared with matched undistorted versions, because such paired images were not available. T2-weighted images therefore served as anatomical references. Mask overlap between corrected diffusion-weighted images and T2-derived prostate masks was used to assess anatomical agreement. DeDistortNet improved agreement in the whole prostate and showed larger gains in the peripheral zone. Under extreme distortion, peripheral zone overlap improved from 0.46 to 0.79. Under severe distortion, it improved from 0.57 to 0.78.

Radiologist assessment also showed better image quality after correction. In 20 exams, scores for geometric distortion and anatomical delineation improved after DeDistortNet correction. Lower scores indicated better quality. Improvements appeared across the distortion categories, including severe and extreme cases. Reader assessment also included the need for repeat acquisition. The results support the potential of the model to improve prostate boundary visibility, reduce distortion-related image problems and support more reliable review of diffusion-weighted prostate MRI.

DeDistortNet corrects prostate diffusion-weighted MRI distortion by combining image signal information with T2-weighted anatomical guidance. The model does not require extra acquisitions or changes to the imaging protocol. Improvements were strongest in severe and extreme distortion, especially in the peripheral zone. Radiologist assessment also showed better geometric fidelity and clearer prostate boundary delineation. Further validation remains important because training relied on simulated distortion, scanner and protocol variation were not fully tested, generative models may produce unrealistic structures, and the current method works slice by slice rather than across full volumes.

Source: European Radiology Experimental

Image Credit: iStock