Feasibility and accuracy of artificial intelligence for automatic qualification of two-dimensional and three-dimensional left ventricular ejection fraction and global longitudinal strain in echocardiography

Abstract

To investigate if fully automated measurements of two-dimensional (2D) and three dimensional (3D) echocardiographic left ventricular ejection fraction (EF) and 2D global longitudinal strain (GLS) using a novel technology based on machine learning, deep learning and artificial intelligence (AI) are feasible and comparable with a conventional application by echocardiography cardiologists. Methods: From 12/2020 to 12/2021, 208 patients with heart disease with a wide range of EF were enrolled in the study in Vietnam National Heart Institute, Bach Mai hospital. Three standard apical cine-loops were analyzed using the AI model. The AI automated method measured 2D EF, 3D EF and GLS and was compared with conventional methods performed by echocardiography cardiologists. Results: The AI method succeeded to both correctly classify all three standard apical views in 95% of patients (four-chamber view: 94,2%; two-champer view:94,7% và three-chamber view: 95,2%.) and perform precise timing of cardiac events. There were strong correlations between 2D/3D EF measured by AI and the 2D/3D EF measured by echocardiography cardiologists (r = 0,78, p<0,001 and r = 0,65, p<0,001, respectively). There was a strong correlation between GLS measured by AI and the GLS measured by echocardiography cardiologists on speckle tracking echocardiography (r = 0,71, p< 0,001). Conclusion: Fully automated measurements of 2D/3D left ventricular ejection fraction and global longitudinal strain using a novel machine learning, deep learning AI-based technology are feasible and fast and they yield results comparable with conventional methods performed by echocardiography cardiologists.