The exact localization of intrahepatic vessels in relation to a tumour is an important issue in oncological liver surgery. For computer-assisted preoperative planning of surgical procedures high quality vessel models are required. In this work we show how to generate such models on the basis of registered CT or MRI data at different phases of contrast agent propagation. We combine well-established intensity-based rigid and non-rigid registration approaches using Mutual Information (MI) as distance measure with a masking strategy as well as intensity inhomogeneity correction for MRI data. Non-rigid deformations are modeled by multilevel cubic B-splines. Quantitative evaluations of 5 MRI and 5 CT image pairs show that the liver moves rigidly 7.1 (± 4.2) mm on average, while the remaining non-rigid deformations range from 2 to 3 mm. As a result we find that masked rigid registration is necessary and in many cases also sufficient. After non-rigid registration the matching shows no deviations in most cases.