A machine learning solution based on the Structured Random Forest algorithm achieved competitive results compared to the semi-automatic Active Appearance Model for fully automating heart segmentation.
Does a machine learning solution based on the Structured Random Forest algorithm provide competitive segmentation of the left ventricle and myocardium on 2D echocardiographic data compared to the Active Appearance Model?
A machine learning approach using the Structured Random Forest algorithm can fully automate the segmentation of the left ventricle and myocardium on 2D echocardiograms with competitive performance to semi-automatic methods.
2D echocardiography remains nowadays the main clinical imaging modality in daily practice for assessing the cardiac function. This task requires an accurate segmentation of the left ventricle (LV) and myocardium at end diastole (ED) and systole (ES). Because of intrinsic high variability in image quality in ultrasound data, manual interactions are still needed to obtain a precise delineation of the heart structures. This is both time consuming for specialists and not reproducible. In this study, we investigate a machine learning solution based on the Structured Random Forest algorithm to fully automate the segmentation of the myocardium and LV on heterogeneous clinical data. We compare its performance to the semi-automatic state of the art Active Appearance Model (AAM). The competitive results that were achieved lead us to believe that supervised learning may be the key to automatic heart segmentation.
Leclerc et al. (Fri,) conducted a other in Echocardiographic segmentation. Structured Random Forest algorithm vs. Active Appearance Model (AAM) was evaluated on Segmentation performance. A machine learning solution based on the Structured Random Forest algorithm achieved competitive results compared to the semi-automatic Active Appearance Model for fully automating heart segmentation.