Two-dimensional echocardiography (2D echo) is the most widely used cardiac imaging techniques in clinical applications. Boundary delineation of the heart, especially the left-ventricle (LV), is essential to calculate the clinical parameters. Currently, LV segmentation from 2D echo is conducted manually or using semi-automatic techniques. In this study, machine learning techniques were employed. U-net, which is a fully convolutional network, and segAN, which is a generative adversarial network, were trained and evaluated. Training was conducted on the in-house dataset, which consists of 2108 porcine images from 10 different subjects. This dataset was the first dataset, which consists of six standard projections of 2D echo over the entire cardiac cycle. Transfer learning was used for long-axis projections to compensate the limitation of in-house dataset using Cardiac Acquisitions for Multi-structure Ultrasound Segmentation dataset. The models were evaluated on test images by computing metrics such as the dice metric. U-net and segAN models outperformed the level-set method, a traditional segmentation technique. The average dice metric of U-net was 0.903 for LV cavity and 0.787 for LV myocardium. The average dice metric of segAN was 0.912 for LV cavity and 0.801 for LV myocardium. Previous reconstruction algorithm was improved and validated to generate the 3D LV geometry from segmented images. Physiological parameters were calculated from reconstructed geometries with about 15% error which is similar to the previous methods using 2D echo compared to the gold standard MRI. In this study, fully automated algorithm to generate 3D LV geometry from 2D echo images was introduced by combining machine learning segmentation technique and 3D reconstruction algorithms. This algorithm facilitates the patient-specific LV modeling and simulation without expert's knowledge and effort.
