Atlas-based segmentation of pathological knee joints.

Efficiency, comparability and simplicity are key aspects for user acceptance of surgical planning systems in the long term. Automatic segmentation and identification of geometric reference systems of the anatomical structures are essential to fulfill these requirements. A statistical motivated shape atlas of the knee joint, based on 235 normal and abnormal MR and CT volume sets, is constructed for automatic segmentation of CT image data. In the first step of the atlas construction, the bony structures of the knee were segmented semi-automatically and processed into a dense and a sparse triangulated surface mesh to obtain training data sets. To establish an inter-individual correspondence, a skeleton-based registration method is used. The registered sparse surface meshes are retriangulated to estimate a pointwise inter-individual correspondence. The shape atlas is build upon these correspondences and integrated into a segmentation algorithm. An iterative segmentation scheme is proposed, which consists of a combination of the iterative-closest-point algorithm for spatial registration and of a downhill-simplex optimization procedure for deformation of the statistical motivated shape atlas to the image data. We expect the statistical shape model to be a robust and image modality independent method for the segmentation of pathological knee joints in CT image data.