Segmentation of 2D Echocardiography Images using Residual Blocks in U-Net Architectures

In this paper, we analyze the results of automatic contouring of the left cardiac structures using deep learning UNet architectures. We use the Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) fully-annotated dataset which contains of the left inner cardiac regions (left ventricle (LV) and left atrium (LA)) with two and four-chamber views of 500 patients. The public-available dataset contains images of 450 patients which was divided into 9 folds with 50 patients each based on image quality and ejection fraction (EF) value. 8 folds were used for training models while the last one for their validation. Every fold was used as validation one in 9 independent models of each architecture and the final result is the mean Dice value of those models. The best contours identification accuracy by the Dice coefficient of the LV endocardium, LV epicardium and LA on validation subsample among our trained models was obtained with U-Net ResNet34 architecture: 0.9348, 0.9459, 0.9038. Those scores show that U-Net architectures with residual blocks achieve better results than U-Net architectures without additional skip-connections.