A Bayesian Similarity Function for Segmentation using Anatomical, Shape-Based Models

FMRIB Technical Report TR05MJ1

Mark Jenkinson

Oxford Centre for Functional Magnetic Resonance Imaging of the Brain (FMRIB),
Department of Clinical Neurology, University of Oxford, John Radcliffe Hospital,
Headley Way, Headington, Oxford, UK

Abstract:

Shape-based segmentation involves fitting a flexible model of anatomical shape to a measured image. It is important to be able to utilise probabilistic prior information about shape, and to combine this with data-driven likelihoods. This is naturally achieved within the Bayesian framework. In this paper a probabilistic similarity function for the anatomical, shape-based segmentation problem is derived using a fully Bayesian approach. Furthermore, it incorporates prior, probabilistic information without the need for additional ad-hoc parameters. Preliminary results show that this similarity function is more robust and accurate than simpler versions based on the same image formation model.