Abstract This paper presents ongoing research on a semi-automatic method for computing, from CT and MR data, patient-specific anatomical models used in surgical simulation. Up to now, virtual anatomical models for surgical simulation have been highly task-specific and obtained by highly elaborated computer–user interaction, including segmentation and meshing. We propose a minimally supervised procedure for extracting from a set of MR and CT scans a highly descriptive anatomical mesh, leading not merely to one generic model, but to a family of patient-specific models. This procedure integrates a distortion-tolerant mutual information MR-CT registration, a new tissue classification exploiting global spatial cues, a simplex-based surface meshing model to identify and triangulate the relevant anatomical boundaries, and a final almost-regular tetrahedralization of tissue volumes surrounded by these boundaries. This procedure is designed to be extensible to other surgical applications.