Res-DUnet: A small-region attentioned model for cardiac MRI-based right ventricular segmentation

Abstract

Right ventricular function has been associated with a variety of cardiovascular diseases. In the clinical study of right ventricular function, an important step is segmentation of the right ventricle so that functional indicators of the heart can be quantified and evaluated based on the segmented region. Compared to left ventricle, right ventricle (RV) is more difficult to segment due to its irregular shape variations and blurred borders. To improve segmentation accuracy on slices with very small regions of RV, a two-stage deep learning model called Res-DUnet is developed to segment the right ventricle on short-axis slices of cardiac magnetic resonance imaging. The model is divided into two modules, the localization module for extracting the regions of interest (ROIs) and the segmentation module for segmenting the right ventricle. The two-stage model achieves a mean dice score (DSC) of 90.55, a mean Hausdorff distance (HD) of 6.67 and a mean recall of 92.68, outperforming the state-of-the-art models. The clinical indicators derived from the model are analyzed for consistency with the ground truth. The results showed that the model’s performance is comparable to that of the radiologists.