In Statistical Atlases and Deformable Registration (Task 2.3), we have completed the statistical atlas of prostate cancer distribution to include 130 datasets, which our Georgetown University and DoD/CPDR collaborators have provided. We have made arrangements for an additional 158 datasets. We have extended our 2D TRUS segmentation method to 3D ultrasound, and we have applied this method successfully for automated segmentation and atlas registration. We have significantly extended our work on automated FEM generation, which is the cornerstone for performing biomechanical simulations of soft tissue deformations and which will be used to train statistical shape models. We have received an NIH R01 grant to perform a clinical validation study using our statistical atlas of prostate cancer and clinical trial has began at BWH. Feasibility studies show very promising results in applying our automated prostate segmentation and atlas warping method to intra-operative MR images from the BWH. We have started to develop a new approach to estimating high-dimensionality probability density functions that capture deformations of anatomical structures in a statistical fashion. This new approach is based on best basis selection methods from over-complete libraries of wavelet packet functions. It transcends significant limitations of more traditional models, such as principal component analysis, which arise when high-dimensionality deformations are to be statistically modeled using relatively few samples. Finally, joint work between Dr. Taylor's and Dr. Davatzikos' laboratories during the past years has converged in common mesh generation and deformable statistical atlases of the pelvis and prostate.

We will continue along four lines of research. The heavy computational burden of the automated ultrasound segmentor/warper is currently a limitation. We will work on significantly shortening processing time by: 1) determining the system's parameters that are essential for accurate registration of a model boundary with a patient's images, and possibly allowing the system to input limited expert-drawn knowledge; and 2) taking advantage of the fact that tissue deformation during the procedure will be relatively small and, therefore, can be followed in real-time using a tracking method initiated at the beginning of the process. Initiation of the model will require a great deal of computation, but it is performed only once. Second, we also continue our work on biomechanically simulating soft tissue deformations, in order to train a statistical shape model that can be used as a prior shape in real time to track soft tissue deformations from limited intra-operative data (e.g. fluoroscopic or ultrasound images). We have recently developed automatic remeshing methods, so that we can handle large deformations while maintaining quality of the Finite Element mesh. Third, we will combine biomechanical simulations and real experimental data with our statistical model described in the previous paragraph, in order to obtain fast statistical models that estimate and track anatomical deformations during a variety of medical procedures. Finally, we will augment our mesh generation, deformation, and atlas construction procedures for the pelvis and knee atlases, so that image intensity variations, which might reflect variations in bone density, are fully exploited and analyzed.