In this paper, a method for deformable registration of male pelvises in CT images is proposed. It is composed of three major steps. Firstly, an attribute vector is defined for each voxel in the image, and is used to reflect the density distribution of the bone tissues in the neighborhood of that voxel. Different anatomical structures thus own different attribute vectors. Secondly, the match of two images under a particular deformation is evaluated by integrating the similarities of attribute vectors on the corresponding voxels. Finally, a set of hierarchical deformation strategies are employed, in order to avoid the local minimum match of two images. The performance of this deformable registration method has been demonstrated by both synthesized and real CT images of pelvises.