An MRF-based statistical deformation model for morphological image analysis

Abstract

As collections of 2D/3D images continue to grow, interest in effective ways to use the statistical morphological properties of a group of images to enhance biomedical image analysis has surged. During the last several years, advances in non-linear registration techniques have made possible the fast estimation of highly accurate deformation fields with dense feature correspondences between two images. Recently, statistical deformation models (SDMs) have emerged as effective methods to capture the statistical and structural properties of a collection of images directly from a set of deformation fields. We present a method to create a robust SDM model that can be used in multiple biomedical applications including image classification, diagnosis, generation, and completion. In particular, we introduce a Markov-based SDM model which uses the deformation properties and contextual relationships to more effectively learn the statistical morphological properties of a group of images. To show the strengths and limitations of our approach, the framework has been tested with synthetic and real-world medical volumes.