Generation and application of statistical shape models for the segmentation of abdominal organs in radiotherapy planning
Michael Quicken · Repository for Publications and Research Data (ETH Zurich) · 2000
This thesis addresses the problem of identifying anatomical structures in volumetric image data resulting from medical imaging.In particular the interest lies in the segmentation of abdominal structures in image volumes acquired by computer tomography modalities.This results from the driving application in radiotherapy.Treatment planning for irradiation of malignant neoplasms requires computer tomography images due to the similar physical behavior of the radiation used for imaging and for treatment.A frequent inelication for raeliotherapy is prostate cancer, which motivated the orientation on segmentation of abdominal structures in male patients.The employed methods, however, are of more general applicability.Computer tomography provides only images of limited contrast for the structures of interest, ruling out rnany approaches to segmentation for this application.The segmentation framewerk presented makes use of three-dimensional statistical anatomical models.A large set of computer tomography images of the male abdomen has been gathered during clinical routine and segmented manually.From these data sets shape and appearance models of anatomical structures, especially the bladder and the prostate, have been createel.Fitting these models to new patient data sets provides segmentations of the structures.Special attention is payed to parameterization of surface meshes, a vital step for the transition from binary volumes to parametric surfacc clescriptions.This transition is necessary as binary volumes result from manual segmentation, but parametric surface descriptions are usecl for model buileling.Former techniques were not able to handle the size of the objects that are considered here, necessitating the development of a new hierarchical technique.The framewerk was incorporated into different approaches to model fitting based on limitation of shape variance by surface approximation and on the optimization of a statistically justified goal function.Multiple shapes have been treated together for shape moeleling and the results from model fitting have been compared to results from fitting of single objects.