Segmentation of Coronary Arteries using Radial Basis Function Neural-Network

Biplane and digital subtraction angiography (DSA) have brought about important advances in the diagnosis and treatment of cardiovascular anomalies by allowing for blood flow measurements, estimation of the regional wall stress and study of myocardium motion. Segmentation of the coronary arteries is a critical first step towards an automated interpretation of angiographs. We present an analysis of neural network methods based on a Radial Basis Function (RBF) and back-propagation (BP) network applied to segmentation of the coronary arterial tree. The results of the neural network based segmentation are compared with segmentation techniques based on a delineation algorithm. Features like vessel diameter and centerline coordinates are extracted for segmented images and compared for the various segmentation methods. The network methods are based on first evaluating the best number of cluster partitions and then automatically obtaining the vectors for training. The pixel gray-level values in a small neighborhood along with information of ridges are utilized to provide the training vectors. The ridge locations indicate high likelihood of continuous points on the artery. A discussion of the learning and generalization characteristics for segmentation, by the networks, is presented for multi-view DSA images and tube phantom simulations.