2D Echocardiography Image Segmentation via Patch-Based Generative Adversarial Network

Accurate left ventricle (LV ) segmentation remains challenging due to the complex anatomical variability and imaging artifacts present in echocardiograms. Our ultimate objective is to develop a robust segmentation method for accurate segmentation of the LV , thereby enhancing the estimation of clinical indices. In this study, we introduce a novel hybrid generative adversarial network (GAN) architecture called patch-based GAN for cardiac image segmentation. Our contribution involves using two backbone models: UNET and a patch-based neural network (PBNN). UNET is employed for precise feature extraction and spatial preservation, aiding in constructing the latent space for the generator (G). Meanwhile, PBNN enhances the discriminator’s ability to accurately distinguish between the generated mask and the ground truth mask. The proposed patch-based GAN framework allows the network to learn from both the input data and adversarial feedback, improving its ability to handle variations and artifacts in echocardiograms. The experiment is conducted on the CAMUS dataset, comprising an echocardiographic dataset from 500 patients in 2-chamber (2CH) and 4-chamber (4CH) views at end-diastolic (ED) and end-systolic (ES) phases. The proposed method achieves high accuracy in LV endocardium segmentation, with mean Dice values of 0.96 and 0.94 for the 2CH view and 0.959 and 0.953 for the 4CH view at ED and ES phases.