The fully automatic extraction of anatomical structures from medical image data is an ill-posed problem. Often it cannot be solved satisfactorily without taking additional high level information about the shape of the particular structure into account. Since the early 1990s, the so-called Statistical Shape Models (SSMs) have arisen as a powerful means in order to integrate this shape information into the segmentation process. Statistical shape models are used to describe the shape variability that resides within a particular object class. The needed information is thereby extracted from a set of hand-labeled training shapes. In the majority of shape models, the variability inside the object class is represented by global linear combinations of several training shapes. Yet, a widespread problem is that the number of hand-labeled training shapes is often insufficient in order to model the variability of complex object classes. This is because generating hand-segmented reference data is a tedious and time-consuming task. Common approaches to cope with this problem are to partition the shapes and to model the shape variability independently across the predefined segments or to allow artificial shape variations which cannot be explained through the training data. However, both approaches have their drawbacks. In this thesis, we propose another approach to handle the problem of limited training data without the need for any predefined segments and without allowing shape variations which cannot be explained through the training data. We call our approach the Locally Deformable Statistical Shape Model (LDSSM). Our idea is to allow a unique solution in each element of the underlying data domain and to couple the local solutions via smoothness constraints. This new formulation allows us to model complex object classes with only a few training shapes at hand. Furthermore, we provide a sound mathematical foundation in order to embed our new LDSSM as a shape prior into the well-known variational image segmentation framework, and we show how it can be used to seamlessly adapt from a globally shape constrained segmentation result to a locally shape constrained segmentation result. We demonstrate the excellent performance of our new approach in several two-dimensional as well as three-dimensional image segmentation and shape reconstruction tasks.
From Global to Local Statistical Shape Priors - Novel methods to obtain accurate reconstruction results with a limited amount of training shapes
Von globalem zu lokalem statistischen Formvorwissen - Neue Ansätze für die Berechnung exakter Rekonstruktionsergebnisse auf Basis einer begrenzten Anzahl an Trainingsdaten
2016
Sonstige
Elektronische Ressource
Englisch
Single and sparse view 3D reconstruction by learning shape priors
British Library Online Contents | 2011
|Visual Saliency with Statistical Priors
British Library Online Contents | 2014
|Extraction of left ventricle borders with local and global priors from echocardiograms
British Library Online Contents | 2013
|