Integrating a gradient–based difference operator with machine learning techniques in right heart segmentation

Abstract

In this research a three step method for right heart segmentation based on a gradient– based difference operator and machine learning techniques is reported. The proposed method is applied to human heart multi–slice computerized tomography (MSCT) volumes. The first step is the preprocessing, where a gradient–based difference operator is applied to exploit the functional relationship between the original input image and its edge enhanced version. In the second step, the least squares support vector machines (LSSVM) are used with a double purpose. First, an appropriate volume-of-interest is automatically established in order to isolate the structure to segment. Second, another LSSVM is trained for locating the voxels required for initializing the seed based clustering procedure. In the third step (segmentation step), the preprocessed volumes are subsequently processed with an unsupervised clustering technique based on simple linkage region growing. Dice score is used as a metric function to compare the segmentations obtained using the proposed method with respect to ground truth volumes traced by a cardiologist. The right atrium, pulmonary valve, right ventricle and venae cavae are segmented from 80 cardiac MSCT volumes. Reported metrics confirm that this method is a promising technique for right heart segmentation.