Relational statistical deformation models for morphological image analysis and classification

Abstract

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 of the morphological statistical properties of a group of images has proven to be extremely important in medical imaging. This paper introduces Relational Statistical Deformation Models, or RSDMs, as a generic modeling technique to capture the morphological statistical deformation properties of a collection of images. Deformation fields, such as those obtained from non-linear registration techniques, are used to learn the morphological properties of a group of images and to train a statistical model capable of solving multiple imaging tasks such as image classification. To compensate for noise and registration errors, RSDM treats each local deformation as a random variable and builds the statistical model as a Markov Random Field (MRF). Once an RSDM model has been created, the same model can be used to solve multiple imaging tasks such as image classification, diagnosis, generation, and denoising. The focus of this paper is to introduce RSDM models and illustrate their effectiveness in image classification tasks. To show the advantages and limitations of RSDMs, a collection of brain MR images was used to create a model to automatically identify subjects with Alzheimer's.