New ideas for the extraction of finite element (FE) models from high-resolution computed tomography datasets are presented. The multi-step approach starts with a 3D region-growing algorithm in order to extract the outer voxel based iso-surface. This information is used to compute a voxel model of the cortical shell. The next step provides triangulated surfaces of the outer bone contour. Three-dimensional deformable models using a gradient vector flow and a multi-level mesh resampling are used. These meshes are self-regularising and of high quality. A further step contains a new self correcting cortical shell thickness evaluation algorithm, which results in topological conform smooth inner and outer compact bone iso-surface meshes. Such iso-surfaces can be used for numerically efficient FE models, which are of bio-mechanical and clinical importance. Details of the approach are described and applications with respect to a human proximal femur and vertebral body are shown.