The best achievable accuracy of the algorithm is evaluated on simulated projections of a virtual heart phantom, showing that it is capable of extracting coronary centerlines with an accuracy that is mainly limited by projection and volume quantization (0.25 mm). The algorithm is reasonably insensitive to residual motion, which means that it is able to cope with inconsistencies within the projection data set caused by limited gating accuracy and respiration. Its accuracy on clinical data is evaluated based on expert ratings of extracted models of 17 consecutive clinical cases (10 LCA, 7 RCA). A success rate of 93.5 % (i.e. with no or slight deviations) is achieved compared to 58.8 % success rate of semi-automatically extracted models.