Learning MRI Artifact Removal With Unpaired Data

Retrospective artefact correction (RAC) improves image quality post acquisition and enhances image usability. Recent machine-learning-driven techniques for RAC are predominantly based on supervised learning, so practical utility can be limited as data with paired artefact-free and artefact-corrupted images are typically insufficient or even non-existent. Here we show that unwanted image artefacts can be disentangled and removed from an image via an RAC neural network learned with unpaired data. This implies that our method does not require matching artefact-corrupted data to be either collected via acquisition or generated via simulation. Experimental results demonstrate that our method is remarkably effective in removing artefacts and retaining anatomical details in images with different contrasts. To remove artefacts from medical imaging, machine learning can be a useful tool, but supervised approaches need examples of the same image with and without artefacts. Liu et al. present a method to train an artefact removal network without needing matching images of corrupted and uncorrupted images.