Nonparametric Joint Shape Learning for Customized Shape Modeling
Computerized Medical Imaging and Graphics, 34(4), 298-307, 2010
in virtual prototyping of implants and anatomical modeling of organs. Some applications in customized prototyping are realized for providing patient-specific models for bones, teeth, and hearing aid devices. Estimating geometry of organs before, during, and after surgery and modeling bilateral relationships between the symmetric structures in the brain are other applications. In problems of customized shape design, the first step involves acquiring a rough raw data of the structure to be modeled. For instance, impressions of the teeth or the ear canal can be acquired via a laser scan; or 3- dimensional (3D) computed tomography (CT), magnetic resonance (MR) image volumes of the patient are obtained and 3D models of the organs are extracted from these images. In the next step, certain operations and rules are applied to the model with constraints from the input patient data, which is at this point an unprocessed 3D surface model. If one would like to predict the intra-operative shape of a particular organ, for instance prostate, a natural approach would be to acquire a pre-operative CT or MR scan of the patient and extract the prostate surface, and apply a learned transformation, to the pre-operative surface to generate the intra-operative geometry of the prostate. A similar problem arises in hearing aid design: one obtains an impression of the ear geometry, and estimates the final hearing aid device, which should comfortably fit to a patient’s ear as well as satisfy other constraints in terms of the electronics components, ventilation vents, and so on. An example problem is depicted in Fig. 1 for illustration, where a complete tooth model is shown along with separate root and crown parts: in dental implants, the decayed or fractured crown part of the tooth is covered with a prosthetic crown, whose shape should be customized for a good fit. We achieve a solution to such problems via a shape optimization method that jointly models multiple shapes either consisting of different parts of one structure or different phases of a given structure.