Echocardiographic segmentation based on semi-supervised deep learning with attention mechanism

Echocardiographic examination is one of the main methods for clinical diagnosis, management and follow-up of heart diseases. Echocardiographic segmentation is an essential step for obtaining precise measurements and accurate diagnosis. However, the current methods are mostly time-consuming, relatively subjective, and produce inconsistent results due to varying ultrasound image quality. To solve these problems, we propose an automatic 2D echocardiographic segmentation method, which is objective and robust for the change of image quality. Specifically, our method first constructs an echocardiographic motion estimation network to extract the heart motion features for the echocardiographic segmentation network. Then, based on semi-supervised learning, the echocardiographic segmentation network is trained by labeled images’ ground truth and unlabeled images’ pseudo labels, which are derived from the motion features. In addition, we introduce attention mechanism to observe its impact on segmentation performance. Experimental results show that the peak signal-to-noise ratio and the structural similarity index between the target images and the images reconstructed by the motion features are over 30dB and 92%, respectively. The echocardiographic segmentation network achieves 95.93% accuracy and 90.94% dice similarity coefficient in the segmentation of cardiac end-diastolic, and achieves 96.06% accuracy and 91.51% dice similarity coefficient in the segmentation of cardiac end-systolic. These results prove that the motion features and segmentation results obtained from our method are effective and accurate. Our code is publicly available at: https://github.com/cherish-fere/motion_net