A Nearly-Automated Approach for Left Ventricular Segmentation using Feature Asymmetry from Real-time 3D Echocardiography

Quantification of left ventricular (LV) function from real-time 3D echocardiography has already been proved to be more accurate compared to LV function derived from 2D echocardiographic imaging. However, tools available in clinical practice depend on cardiologist's experience in selecting the correct views for volume computation, tracing endocardial contours or selecting some feature points. The aim of this study was to develop a new LV segmentation algorithm driven by feature asymmetry (FA) information. The proposed approach was based on a 3D modified Malladi-Sethian level-set model, driven by FA as an edge indicator, followed by a curvature flow and isosurface extraction. It was tested and validated on 3D echo data made available for the challenge on endocardial 3D ultrasound segmentation, organized during MICCAI 2014, in 9 patients, by correlation and Bland-Altman analysis for end-diastolic and end-systolic volumes and ejection fraction. LV surfaces were compared by means of absolute and Hausdorff distance and Dice coefficient. Clinical indexes derived from the detected surfaces were in good agreement with the reference ones. These promising results were supported by the results obtained by comparing LV surfaces. This study provides the basis for a fast and accurate quantification of cardiac function from real-time 3D echocardiography.