Registration & Modeling of Shapes with Uncertainties: Contributions and Applications to Knowledge Based Segmentation
Maxime Taron · 2007
In the recent years, we have witnessed a revolution on new non-invasive means for human and biological tissues imaging. The use of computer aided-techniques has emerged naturally as an efficient pre-screening and post treatment evaluation procedure. This often involves mathematical modeling of organs which usually refers to the three following steps: (i) determine/extract the structure of interest, (ii) provide a mathematical model to describe these structures and (iii) estimate the variations of the parameters in the proposed model. In this thesis we propose novel means to address the above tasks. The modeling of organs is performed with the description of the deformations. It therefore involves shape registration with the definition of a reference shape and a deformations space. In this thesis we introduce the use of uncertainties in the registration, defined as covariance matrices in the deformations space which indicate the amount of confidence in the obtained registration. The next step consists in modeling the shape variations based on a training set representing various instances of the organ under study. In this thesis we extend the state of the art that does not account for registration errors and introduce a method that propagates registration uncertainty to the modeling step. The last contribution of the thesis is in the area of knowledge based segmentation and consists of introducing a segmentation-by-deformation approach where the use of uncertainties both in the model as well as image space are considered. Segmentation of the cardiac left ventricle on CT scan and of the corpus callosum on MR-images using the above mentioned-methods are considered as applications to demonstrate the extreme potentials of our approach.