Artifact Detection in Cardiac MRI Data by Deep Learning Methods

Abstract

Image acquisition procedure of cardiac MRI may not always result in desirable image quality due to patient movement during the scan, the inability of the MR machine to focus on the appropriate region or the patient's arrhythmia. This study focuses on detecting motion artifacts on cardiac MRI short-axis scans while analysing the effect of handling the data in different shapes. In this regard, two models are developed using deep learning methods. The former processes the data as independent 2-D slices using convolutional neural networks, whereas the latter combines convolutional neural networks with recurrent neural networks to take temporal information into account. Performance has been reported on 200 cardiac MRI short-axis view samples by using 10-fold cross-validation. Numerical results demonstrate that the former network provides higher detection rates than the latter, particularly 0.87 and 0.92 area under curve (AUC) score for 2-D and 3-D models, respectively. We owe this significant gap in performance to have more samples in favour of the 2-D model.