Integrating Statistical Models of Bone Density into Shape Based 2D-3D Registration Framework

We present a framework to deformably register simulated X-ray images to a combined statistical model of pelvis anatomical structure, created from a population of CT scans. The primary contributions are: 1) a framework to create and analyze bone density variations, separate from shape variations and 2) an augmented 2D/3D registration framework that couples shape and density priors to create accurate patient specific models. Our statistical model representation consists of a tetrahedral mesh for approximating bone shape and Bernstein polynomials defined within each tetrahedron for bone density. All datasets in the given population are registered deformably to a template CT dataset. The shape and density statistics are extracted using principal component analysis on the corresponding mesh vertices and voxels of the shape-free deformed subjects respectively. In the registration framework, we register the 2D input images to the 3D shape prior and estimate the bone density parameters in a least-squares like setup by projecting the density modes on to the input image space. This approach was tested using leave-n-out experiments, with n = 8, datasets using an atlas of 63 full pelvis CT datasets.