In this paper, we present a novel approach to rigidly register intraoperative electromagnetically tracked ultrasound (US) with pre-operative contrast-enhanced magnetic resonance (MR) images. The clinical rationale for this work is to allow accurate needle placement during thermal ablations of liver metastases using multimodal imaging. We adopt a model-based approach that rigidly matches segmented liver surface shapes obtained from the multimodal image volumes. Towards this end, a shape-constrained deformable surface model combining the strengths of both deformable and active shape models is used to segment the liver surface from the MR scan. It incorporates a priori shape information while external forces guide the deformation and adapts the model to a target structure. The liver boundary is extracted from US by merging a dynamic region-growing method with a graph-based segmentation framework anchored on adaptive priors of neighboring surface points. Registration is performed with a weighted ICP algorithm with a physiological penalizing term. The MR segmentation model was trained with 30 datasets and validated on a separate cohort of 10 patients with corresponding ground truth. The accuracy and robustness of the method were assessed by registering four US/MR datasets, yielding accurate landmark registration errors (3.7 ± 0.69mm) and high robustness, and is thus acceptable for radiofrequency clinical applications.