The learning of the shape and appearance behavior of complex anatomical structures is of growing importance in the successful use of medical imaging data. We propose a method to simultaneously learn a model of shape variation and the behavioral structure of objects in volumetric data sets. The algorithm performs a group-wise registration of a set of examples, and accounts for the heterogeneous deformation or variability properties of the data. We use the method for the in situ analysis of endograft deformation in the thoracic aorta during the cardiac cycle. The method is based on an emerging model of the shape variation, which is learned autonomously from a gated computed tomography sequence. It automatically adapts to the highly non-uniform elasticity properties of the structure during learning. The resulting deformation model is used for the measurement of global and local characteristics of the endograft movement. The method allows for the in situ localization of the stent during the cardiac cycle, and the measurement of its deformation. Furthermore, it makes the comparison of different endograft designs possible, and can serve as a basis for fitting a physical model of the endograft- and vessel surface to individual patients. The latter is essential for long-term risk assessment of the impact of endografts in highly mobile areas. We evaluate the approach on 10 data sets from patients that underwent endograft placement after traumatic ruptures of the thoracic aorta.