A new paradigm for the characterization of structure appearance is proposed, based on a combination of gray-level MRI intensity data and a shape descriptor derived from a priori principal components analysis of 3D deformation vector fields. Generated without external intervention, it extends into 3D more classic, 2D manual landmark-based shape models. Application of this novel concept led to a method for the segmentation of medial temporal lobe structures from brain magnetic resonance images. The strategy employed for segmentation aims at synthesizing, using the appearance model, a deformation field that maps a new volume onto a reference target. Any information defined on the reference can then be propagated back on the new volume instance, thereby achieving segmentation. The proposed method was tested on a data set of 80 normal subjects and compared against manual segmentation as well as automated segmentation results from ANIMAL, a nonlinear registration and segmentation technique. Experimental results demonstrated the robustness and flexibility of the new method. Segmentation accuracy, measured by overlap statistics, is marginally lower (< 2%) than ANIMAL, while processing time is six times faster. Finally, the applicability of this concept toward shape deformation analysis is presented.