Distortion correction of diffusion weighted MRI without reverse phase-encoding scans or field-maps

Summary

A second-generation Synb0 pipeline uses three-dimensional deep learning to create an undistorted diffusion reference from available MRI data. It supports distortion correction when reverse phase-encoding scans or field maps are missing, extending practical preprocessing options for historical and heterogeneous datasets.

Study overview

The study develops a three-dimensional U-net approach to synthesize an undistorted b0 image with structural T1-weighted geometry and diffusion-like contrast. A heterogeneous training dataset was used to improve applicability across acquisition conditions, and performance was evaluated on withheld data against standard reversed phase-encoding correction and intensity-based registration. The synthesized target enabled correction of geometric distortions without collecting an additional field map or opposite-encoding scan. This extends the earlier Synb0-DisCo method and reduces reliance on computationally intensive registration alone. Because the target is synthesized, unusual anatomy, contrast, or acquisition conditions still require quality control; successful correction in the test set is not a guarantee for every new scan.

Main findings

• The pipeline used three-dimensional U-nets and heterogeneous training acquisitions to synthesize an undistorted b0 image.
• Withheld test data showed successful geometric distortion correction without additional correction scans.
• The proposed approach was quantitatively compared with intensity-based registration and FSL TOPUP.