A new methodology for three-dimensional (3D) non-linear registration of articulate objects is presented. This approach is described in the context of a program to screen for mutant mouse phenotypes based on high resolution MR images of whole animals. The screening process requires statistical definition of the normal (non-mutant) mouse anatomy and its variation. The registration algorithm is designed to remove postural differences between mice and thus enable meaningful comparisons between images. The main focus of our approach is anatomical content and its preservation under deformations. Prior knowledge guides the registration through an evolutionary process based on a hierarchical tree of mouse anatomy.