Cross-domain Artefact Correction of Cardiac MRI

Artefacts constitute a paramount issue in medical imaging where the prevalence of artefacts may severely impact the clinical diagnosis accuracy. Specifically, the mistriggering and motion family of artefacts during the cardiac MR image acquisition would eventually damage the visibility of certain tissues, such as the left ventricular, right ventricular, and myocardium. This would cause the ejection fraction to be incorrectly estimated and the patient’s heart condition to be incorrectly evaluated. Despite much research on medical image reconstruction, relatively little work has been done for cardiac MRI artefact correction. In this work, inspired by the image reconstruction literature, we propose to use an auto-encoder-guided mistriggering artefact correction method, which not only corrects the artefacts in the image domain but also in the k-space domain with the introduction of an enhanced structure. We conduct a variety of experiments on photos and medical images to compare the performances of different network architectures under mistriggering artefacts and gaussian noise. We demonstrate the superiority of the cross-domain network in the case of k-space-related artefacts.