Automatic Segmentation of 2-D Echocardiography Ultrasound Images by Means of Generative Adversarial Network

Automated cardiac segmentation from 2-D echocardiographic images is a crucial step toward improving clinical diagnosis. Anatomical heterogeneity and inherent noise, however, present technical challenges and lower segmentation accuracy. The objective of this study is to propose a method for the automatic segmentation of the ventricular endocardium, the myocardium, and the left atrium (LA), in order to accurately determine clinical indices. Specifically, we suggest using the recently introduced pixel-to-pixel generative adversarial network (Pix2Pix GAN) model for accurate segmentation. To accomplish this, we integrate the backbone PatchGAN model for the discriminator and the UNET for the generator, for building the Pix2Pix GAN. The resulting model produces precisely segmented images because of UNET's capability for precise segmentation and PatchGAN's capability for fine-grained discrimination. For the experimental validation, we use the cardiac acquisitions for multistructure ultrasound segmentation (CAMUS) dataset, which consists of echocardiographic images from 500 patients in two-chamber (2CH) and four-chamber (4CH) views at the end-diastolic (ED) and end-systolic (ES) phases. Similar to state-of-the-art studies on the same dataset, we followed the same train-test splits. Our results demonstrate that the proposed generative adversarial network (GAN)-based technique improves segmentation performance for clinical and geometrical parameters compared with the state-of-the-art methods. More precisely, throughout the ED and ES phases, the mean Dice values for the left ventricular endocardium (LVendo) reached 0.961 and 0.930 for 2CH, and 0.959 and 0.950 for 4CH, respectively. Furthermore, the average ejection fraction (EF) correlation and mean absolute error (MAE) obtained were 0.95 and 3.2 mL for 2CH, and 0.98 and 2.1 mL for 4CH, outperforming the state-of-the-art results