Right Ventricle Segmentation via Registration and Multi-input Modalities in Cardiac Magnetic Resonance Imaging from Multi-disease, Multi-view and Multi-center

Quantitative assessment of cardiac function requires accurate segmentation of cardiac structures. Convolutional Neural Networks (CNNs) have achieved immense success in automatic segmentation in cardiac magnetic resonance imaging (cMRI) given sufficient training data. However, the performance of CNN models greatly degrade when the testing data is from different vendors or different centers. In this paper, we introduce the use of image registration to propagate annotation masks from labeled images to unlabeled images as to enlarge the training dataset. Furthermore, we investigated various input modalities including 3D volume, single-channel 2D image, multi-channel 2D image constructed from spatial and temporal stack to extract more features to improve domain generalization in cMRI segmentation. We evaluated our method in M&Ms-2 challenge testing data (https://www.ub.edu/mnms-2/ ), achieving averaged Dice scores of 0.925, 0.919 and Hausdorff Distance of 10.587 mm, 6.045 mm in right ventricular segmentation in short-axis view and long-axis view respectively.