Automatic coronary extraction by supervised detection and shape matching

Automatic coronary extraction has great clinical importance in the effective handling and visualization of large amounts of 3D data. Despite tremendous previous research, coronary extraction remains difficult. Two such difficulties are extraction of both normal and abnormal vessels and reconstruction of exact tree structures based on anatomical knowledge. To solve the first difficulty, we propose a method to learn a classifier of a tubular 3D object with a dimension reduction approach using Hessian analysis. This enables detection of vessel candidate points despite variations in their appearances. Regarding the second difficulty, we propose an approach to apply the MRF framework for vascular structure segmentation. A novelty of the approach is incorporating constraints to avoid topological inconsistency. Correspondences between the candidate points and model points are found using a graph matching process during which, tree structures as per the shape model are simultaneously reconstructed. Experimental results show robustness of the method. The proposed method can improve clinical workflow.