The goals of this proposal are three-fold: 1) acquire realistic, multimodal data sets that can be used for validating quantitative segmentation algorithms, 2) test segmentation algorithms developed under funding from existing grant (IROICA52709), and 3) improve the state of the art in registration through further development of homologous feature definition tools. Realistic validation of segmentation algorithms on data sets with corresponding functional/histological correlation is very important. While validations usually include imaging relatively simplistic phantoms usually constructed from milled shapes in Lucite with doped, water-filled voids, such phantoms grossly under estimate artifacts encountered in routine clinical imaging. We wish to acquire data sets that more closely mimic the clinical setting, but in a setting where ground truth can be known. Imaging rat brains injected with 9L carcinomas, followed by sacrifice and multislice autoradiography which will be reconstructed to mimic FDG PET imaging, will provide multimodal data sets containing both anatomical and functional information. In addition each data set will have at least one correlative histopathologic slide through midlesion where regions of normal tissue, edema, tumor, and necrotic core will be outlined by a pathologist. These data sets will be made available for easy access to the national/international research community. Segmentation algorithms currently under development in our lab through funding from NIH IROICA52709 will be tested using the acquired data sets. These algorithms are more fully described in the preliminary results section of this application. Reported results of testing will include scatter charts of algorithm-computed volumes versus expert-measured volumes, as well as the more sensitive metric of percent non-intersecting volumes between automatic algorithms and expert-defined surfaces. Multivariate segmentation and multimodality registration are intimately connected topics. The accuracy of multivariate segmentation can be greatly enhanced by the addition of registered data sets containing independent information from other modalities, and likewise the registration task can be aided by automatic generation of shape features from an accurate segmentation algorithm. Approaches to registration usually involve either surface correlations, moment matching, or user-defined (usually point) feature matching to estimate the geometric mapping of one data set onto another. Due to the generality of the user-feature identification method, we have developed an affine method to register 3D volume data sets using a weighted least mean square error estimator on features including not only homologous points, but also lines and planes. We wish to continue this work to include use of additional features as well.
