An Automated and Generalizable Technique for Left Ventricle Segmentation in 2D Echocardiography Utilizing Generative Adversarial Network

Accurately segmenting the left ventricle (LV) is crucial for understanding normal heart anatomy and identifying abnormal or diseased structures based on the estimated ejection fraction (EF). A significant challenge in LV segmentation is ensuring generalizability across diverse datasets with various acquisition settings. Addressing this challenge is a key focus, and we aim to provide a solution to ensure a generalizable segmentation. This study utilizes Pix2Pix Generative Adversarial Networks (GANs) for robust and adaptable left ventricle (LV) segmentation in 2D echocardiography images. Specifically, through the integration of adversarial loss, our model learns to generate anatomically accurate segmentations, which helps to enhance accuracy and adaptability across diverse echocardiography datasets. Our evaluation involved two datasets: the Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) dataset and the EchoNet dataset, both captured from a 4-chamber (4CH) view. To assess generalizability, we train the model using the EchoNet training set and test it on the CAMUS data test set, and vice versa. When trained on the EchoNet dataset and tested on CAMUS, the method achieved a mean absolute error of 4.6%, and a correlation coefficient of 0.98 for EF estimation, which outperforms the state of the arts. Conversely, training on CAMUS and testing on EchoNet resulted in a mean absolute error of 9% and a correlation coefficient of 0.75.