Automatic classification of images of an angiography sequence using modified shape context-based spatial pyramid kernels

Coronary angiography is routinely used to screen patients both prior to and during angioplasty. Each angiography study results in a collection of video sequences or “runs” that depict coronary arteries from different viewpoints. A key problem to be addressed in the automatic interpretation of coronary angiography videos is the identification of images depicting coronary arteries in these sequences. In this paper we present a classification approach to distinguish between the coronary arteries and background images using the shape context descriptor and the learning framework of spatial pyramid kernels. Specifically, we extract centerlines of coronary arteries and represent their intensity distributions and layouts using a Mercer kernel formed from the histograms of intensity and shape context. A multi-class support vector machine is then used to classify a new image depicting coronary arteries. Experimental results are presented that show a high degree of accuracy in artery classification using our approach even under variation in appearance due to viewpoint, coronary anatomy differences, disease-specific variations and changes in imaging conditions.