Multi-scale Statistical Models of Shape and Appearance: Application to Modeling Variations in Anatomical Structures

This thesis deals with challenging problems related to the interpretation and quantification of medical images. We develop and apply novel statistical models of shape and appearance to improve the support of various medical practices. We are particularly concerned with two spine-related tasks: the segmentation of vertebral structures in x-ray images, and the analysis of vertebral deformities. These tasks are relevant for various clinical activities such as the diagnosis of spine-related diseases, surgical planning, and studies of disease progression. Within this context, this thesis makes contributions in the following areas:;Second, we present a novel framework for shape analysis and classification of vertebral structures by using pathology-related Wavelet-ICA shape parameters. Within a statistical framework, we are able to establish a statistically relevant relationship between the wavelet-ICA modes of deformations and clinical information. The framework is then extended to achieve multi-vertebrae classification by using Wavelet-ICA parameters. Thanks to the locality of the features, the new classification scheme can assess the condition of several vertebrae simultaneously. Our results show that significant differences are found between groups of healthy and pathological vertebrae.;Third, we develop a novel method for vertebral segmentation by using localized appearance models and a multi-scale shape prior. The new segmentation algorithm incorporates the following: (i) a method for salient point detection in the non sub-sampled Contourlet domain, (ii) Contourlet-based local appearance profiles, and (iii) a multi-scale shape prior to drive the segmentation process. We evaluate the ability of the proposed method to accurately segment normal and pathological vertebral structures.;Finally, we tackle the challenges related to the accurate modeling of the texture and scarcity of training data. We present a new sparse texture appearance model based upon the Contourlet Transform and Independent Component Analysis (ICA). The new model benefits from the non-linear approximation power of Contourlets to achieve high reduction rates. ICA is then employed to capture localized texture variations. We also describe a general framework that integrates our developed shape and texture models in a unified framework for synthesis of photo-realistic pathological x-ray images.;First, in an attempt to overcome the challenges associated with the modeling of high-dimensional shape space, we propose a novel multi-scale statistical shape model. The new model uses concepts of sparsity, best basis selection and Independent Component Analysis (ICA) to extract multi-scale modes of deformations. This method allows for the construction of a localized non-linear shape model with clinically-relevant modes of deformation.