3D RECONSTRUCTION OF CORONARY ARTERIES FROM ROTATIONAL X-RAY ANGIOGRAPHY Ruben Cardenes 1 Alexey Novikov 1 Julian Gunn 2 Rod Hose 2 Alejandro F. Frangi 1,2 1 Center for Computational Imaging & Simulation Technologies in Biomedicine (CISTIB) – Universitat Pompeu Fabra and CIBER-BBN, Barcelona, Spain 2 University of Sheffield, Sheffield, U.K. ABSTRACT This paper presents a 3D reconstruction method of coronary arteries from multiple projections, that works efficiently in two steps: first, a 2D processing is performed to obtain and classify the vessels branches on the projections, estimating also the radii. Then, a novel 3D reconstruction step is per- formed from pair of projections, finding the best intersections between ray bundles corresponding to the same branches to extract the 3D vessel trajectories, that is robust to motion. Topological constraints are included implicitly imposing ves- sel connectivity and tubular construction. The preliminary results presented here from a digital phantom and from real data, show good accuracy and performance of the method. Index Terms— 3D reconstruction, coronary arteries, pro- jections, X-ray angiography, 3DCA, skeletons 1. INTRODUCTION Assessment of coronary artery disease (CAD) is usually car- ried out by Conventional Coronary Angiography (CCA) or Computer Tomography Angiography (CTA), imaging [1], es- pecially multi-detector CT (MDCT), that provides high im- age quality overcoming 2D limitations of CCA such as vessel overlapping and foreshortening. However, rotational X-ray angiography or 3DCA can solve some of the disadvantages of CCA, and compete with MDCT, with the use of motion compensated reconstruction techniques. Additionally, 3DCA can be acquired in the operation room, before and after treat- ment, and still provides higher spatial resolution, making it an extremely useful image modality to analyze coronary ar- teries. The main problem in 3DCA is how to combine the information of several projections compensating the cardiac and respiratory motion, in order to construct a valid vascular model. Additionally, there are some other problems, such as intensity inhomogeneity, vessel foreshortening and overlap of different structures such as the catheter, bones or diaphragm, making this reconstruction a very challenging problem. First author funded by Beatriu de Pinos program, AGAUR, Generali- tat de Catalunya. Work partially funded by CDTI CENIT-cvREMOD grant, Spain. Several techniques have been proposed for this particular application. In [2], the reconstruction is performed computing the 2D vessel centerlines from the projections, and search- ing correspondences in them, to then obtain a 4D field rep- resenting the motion, without using the ECG signal. In [3], the background of the projections is removed using a top hat filter, and then, a thresholded Algebraic Reconstruction Tech- nique (ART) method of [4] is used for ECG gated projections. In [5] a back-projection of the projections’ vesselness into a 3D volume is performed, using all the possible pairs of pro- jections of the same ECG time step.Then, the 3D volume is segmented using fast marching, to finally do a backtracking to obtain the skeleton of the coronaries. In [6] reconstruction is done by an optimization scheme from two uncalibrated an- giographic images with 14 parameters involved, where two errors are minimized, distances and direction vectors of cor- responding skeleton points in the original projections and in reconstructed skeleton projections. The correspondence of centerline points between the two projections is carried out using epipolar constraints. The approach presented here has similar characteristics to the works mentioned above. The method proceeds in two steps: 2D angiographic processing plus a 3D reconstruction stage, as in [2, 3, 5, 6]. However, there is no need here to generate an initial reconstruction from the projections, as car- ried out in [3] and [5], and the novelty of this method is a 3D strategy to obtain the 3D centerlines of the coronaries, using 2D branches correspondences of the 2D skeletons, not point to point as done in [2, 6] . This 3D process is novel, robust, and is able to account for cardiac and respiratory motion. In this paper we will focus in the 3D processing stage. 2. METHOD The two stages proposed are: first, a 2D processing of the projections to identify consistently the arteries’ branches, to classify them across the entire sequence, and to compute the vessel radii. Then, a 3D step is performed to obtain 3D mod- els at different time steps. 618 978-1-4577-1858-8/12/$26.00 ©2012 IEEE ISBI 2012