Structured random forests for myocardium delineation in 3d echocardiography

Abstract

Delineation of myocardium borders from 3D echocardiogra-phy is a critical step for the diagnosis of heart disease. Following the approach of myocardium segmentation as a contour finding task, recent work has shown effective methods to interpret endocardial edge information in the left ventricle. Nevertheless, these methods are still prone to preserve irrelevant edge responses and would struggle to overcome chief ventricle anatomical challenges. In this paper we adapt Structured Random Forests, borrowed from computer vision, for fast and robust myocardium edge detection. This method is evaluated on a dataset composed of short-axis slices from 25 End-Diastolic echocardiography volumes. Results show that the proposed ensemble model outperforms standard intensity-based and local phase-based edge detectors, while removing or significantly suppressing irrelevant edges triggered by ultrasound image artefacts and blood pool anatomical structures.