Generation and visualization of relational statistical deformation models for morphological image analysis

As medical imaging datasets continue to grow, interest in effective ways to analyze the statistical properties and data variability within those datasets has surged. Accurate analysis and understanding of the morphological statistical properties of a group of images has proven to be extremely important in medical imaging. Since the detection of irregular anatomical deformations can increase the ability to identify particular diseases, a number of techniques to capture statistical anatomical properties have been proposed.;Most recent research projects have focused on either statistical anatomical models or visualization systems as vital tools to assist physicians during the diagnosis process. However, innovation and research on statistical models and visualization should be combined to enable accurate analysis of the statistical properties of a group of images. This dissertation focuses on how those two components can be combined to enable flexible analysis of medical images and provide an understanding of anatomical differences, relationships, and variability.;First, we address the problem of statistical analysis through a novel relational model. Relational Statistical Deformation Models, or RSDMs, are introduced as a generic modeling technique to capture the morphological statistical deformation properties of a collection of images. RSDMs take advantage of the information provided by individual deformation fields to build a robust graph-based statistical model which can be applied to multiple image analytics tasks. In general, the morphological framework can be described as a Markov Random Field model which combines a large constellation of probability density functions and uses energy minimization techniques to determine the best solution for different imaging tasks such as image classification and generation. RSDMs have proven to be valuable statistical models in the diagnosis, generation, denoising, and completion of medical imagery. In addition, RSDMs have proven to be effective in the automatic detection of subjects with Alzheimer's disease.;Second, this dissertation advances the field of visualization by introducing four new illustration techniques to effectively summarize statistical deformation properties within a single image. The different illustration techniques can annotate the statistical information obtained from a large group of images into a single model, thus permitting easy exploration of anatomical differences and relationships. Each of the annotation techniques have been applied to synthetic and real-world medical images. In addition, an in-depth user study was conducted to better determine the advantages and limitations of each approach. Results show that there are significant advantages of using statistical illustration techniques over analyzing the set of images individually.