Cardiac MR Image Segmentation and Quality Control in the Presence of Respiratory Motion Artifacts Using Simulated Data

In this work, we propose solutions for the two tasks of the CMRxMotion challenge; 1) quality control and 2) image segmentation in the presence of respiratory motion artifacts. We develop a k-space based motion simulation approach to generate cardiac MR images with respiratory motion artifacts on open-source artifact-free data to handle data scarcity. For task 1, a motion-denoising auto-encoder is trained to reconstruct motion-free images from the pairs of images with and without simulated motion. The encoder part of the auto-encoder is used as a feature extractor for a fully-connected classifier. For task 2, an ensemble of modified 2D nn-Unet models is proposed to tackle different aspects of variations in the data with the purpose of improving the robustness of the model to images hampered by respiratory motion artifacts. All proposed models in this paper are trained using the images with simulated motion artifacts. The proposed quality control model achieves a classification accuracy of 0.75 with the Cohen’s kappa coefficient of 0.64 and the ensemble model obtains the mean Dice scores of 0.922, 0.829, and 0.910 respectively for the left ventricle, myocardium, and right ventricle segmentation on the validation set of the CMRxMotion challenge.