A learning based hierarchical model for vessel segmentation

In this paper we present a learning based method for vessel segmentation in angiographic videos. Vessel segmentation is an important task in medical imaging and has been investigated extensively in the past. Traditional approaches often require pre-processing steps, standard conditions or manually set seed points. Our method is automatic, fast and robust towards noise often seen in low radiation X-ray images. Furthermore, it can be easily trained and used for any kind of tubular structure. We formulate the segmentation task as a hierarchical learning problem over 3 levels: border points, cross-segments and vessel pieces, corresponding to the vessel's position, width and length. Following the marginal space learning paradigm the detection on each level is performed by a learned classifier. We use probabilistic boosting trees with Haar and steerable features. First results of segmenting the vessel which surrounds a guide wire in 200 frames are presented and future additions are discussed.