Heuristic tree searching for pose-independent 3D/2D rigid registration of vessel structures

Abstract

The 3D/2D registration of pre-operative computed tomography angiography (CTA) and intra-operative x-ray angiography (XRA) images in vascular intervention is imperative for guiding surgical instruments and reducing the dosage of toxic contrast agents. In this study, 3D/2D vascular registration is formulated as a search tree problem on the basis of the topological continuity of vessels and the fact that matching can be decomposed into continuous states. In each node of the tree, a closed-solution of 3D/2D transformation is used to obtain the registration results based on the dense correspondences of vessel points, and the results of matching and registration are calculated and recorded. Then, a hand-crafted score that quantifies the qualities of matching and registration of vessels is used, and the remaining problem focuses on finding the highest score in the search tree. An improved heuristic tree search strategy is also proposed to find the best registration. The proposed method is evaluated and compared with four state-of-the-art methods. Experiments on simulated data demonstrate that our method is insensitive to initial pose and robust to noise and deformation. It outperforms other methods in terms of registering real model data and clinical coronary data. In the 3D/2D registration of uninitialized and initialized coronary arteries, the average registration errors are 1.85 and 1.79 mm, respectively. Given that the proposed method is independent of the initial pose, it can be used to navigate vascular intervention for clinical practice.