Ventriculogram segmentation using boosted decision trees

Left ventricular status, reflected in ejection fraction or end systolic volume, is a powerful prognostic indicator in heart disease. Quantitative analysis of these and other parameters from ventriculograms is infrequently performed due to the expense of manual segmentation. We present a method for semi-automatic segmentation of ventriculograms based on a two-stage boosted decision-tree pixel classifier. The classifier determines which pixels are inside the ventricle at key end-diastole and end-systole frames. The classifier is semi-automatic, requiring a user to select 3 points in each frame: the endpoints of the aortic valve and the apex. The classifier uses about 90 feature images, computed from the raw ventriculogram image frames, including simple per pixel gray-level statistics (e.g. median brightness) and image geometry (e.g. coordinates relative to user supplied 3 points). Border pixels are determined from segmented images using dilation and erosion. A curve is then fit to the border pixels, minimizing a penalty function that trades off fidelity to the border pixels with smoothness. Volumes and ejection fraction are estimated from border curves using standard area-length formulas. On independent test data, the differences between automatic and manual volumes (and ejection fractions) are similar in size to the differences between two human observers.