Generalized sparse MRF appearance models

  • Authors:

  • René Donner;Georg Langs;Branislav Mičušik;Horst Bischof

  • Affiliations:

  • Computational Image Analysis and Radiology Lab, Department of Radiology, Medical University of Vienna, Währinger Gurtel 18-22, 1090 Vienna, Austria;Computational Image Analysis and Radiology Lab, Department of Radiology, Medical University of Vienna, Währinger Gurtel 18-22, 1090 Vienna, Austria;George Mason University, 4400 University Drive MS 4A5, Fairfax, VA 22030, USA;Institute for Computer Graphics and Vision, Graz University of Technology, Inffeldgasse 16, 8010 Graz, Austria

  • Venue:
  • Image and Vision Computing
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Image segmentation or registration approaches that rely on a local search paradigm (e.g, Active Appearance Models, Active Contours) require an initialization that provides for considerable overlap or a coarse localization of the object to be segmented or localized. In this paper we propose an approach that does not need such an initialization, but localizes anatomical structures in a global manner by formulating the localization task as the solution of a Markov Random Field (MRF). During search Sparse MRF Appearance Models (SAMs) relate a priori information about the geometric configuration of landmarks and local appearance features to a set of candidate points in the target image. They encode the correspondence probabilities as an MRF, and the search in the target image is equivalent to solving the MRF. The resulting node labels define a mapping of the modeled object (e.g. a sequence of vertebrae) to the target image interest points. The local appearance information is captured by novel symmetry-based interest points and local descriptors derived from Gradient Vector Flow (GVF). Alternatively, arbitrary interest points can be used. Experimental results are reported for two data-sets showing the applicability to complex medical data. The approach does not require initialization and finds the most plausible match of the query structure in the entire image. It provides for precise, reliable and fast localization of the structure.