A Novel Approach for Automated Segmentation of Left Ventricle Based on Bidirectional Myocardium to Endocardium Translation Using Generative Adversarial Network

In echocardiography, accurate segmentation of cardiac structures, particularly left ventricle (LV), is crucial for clinical diagnosis. However, manual segmentation is user-intensive and prone to variability among experts due to intricate anatomical details and imaging artifacts. Our aim is to propose an artificial intelligence (AI)-based technique to automatically segment the LV structure, enhancing accuracy and reducing segmentation time and subjectivity of manual segmentation. We propose a novel Bidirectional generative adversarial network (Bi-GAN) for automated segmentation of LV structures. Specifically, we utilize Bi-GAN for segmenting the endocardium region from available myocardium region and vice versa. The adversarial training, Bi-GAN minimizes the losses and produces the target domain segmentation. The analysis is conducted on the cardiac acquisitions for the multi-structure ultrasound segmentation (CAMUS) dataset, comprising 900 echocardiographic training images and 100 testing images in 4CH views during both end-diastolic (ED) and end-systolic (ES) phases. Results showed a mean Dice of 0.97 for both the ED and ES phases, along with an MAE of 1.90. Conversely, when generating myocardium segmentations from the endocardium, our method attained a mean Dice of 0.88 and MAE of 5.05. Hence, the proposed technique achieves competitive results comparable to the state-of-the-art in LV segmentation, requiring only one set of LV masks (either endocardium or myocardium) as input.