Markov Random Field Based Structure Localisation of Vertebrae for 3D-Segmentation of the Spine in CT Volume Data

David Major
Markov Random Field Based Structure Localisation of Vertebrae for 3D-Segmentation of the Spine in CT Volume Data
[image] [paper]

Information

Abstract

Medical Image Processing is a growing field in medicine and plays an important role in medical decision making. Computer-based segmentation of anatomies in data made by imaging modalities supports clinicians and speeds up their diagnosis making compared to doing it manually. Computed Tomography (CT) is an imaging modality for slice-wise three dimensional reconstruction of the human body in the form of volumetric data which is especially applicable for imaging of bony structures and so for the vertebral column. Most bony structures, such as vertebrae, are characterised by complex shape and texture appearances which turns its segmentation into a difficult task. Model-based segmentation approaches are promising techniques to cope with variations in form and texture of the anatomy of interest. This is done by incorporating information about shape and texture appearance gained from an imaging modality in a model. The model can then be applied to segment the object of interest in target data, however most of the model-based approaches need a model intialisation for a fast and reliable segmentation of the object of interest. This thesis was motivated by novel works on fast anatomical structure localisation with Markov Random Fields (MRFs) and focuses on the sparse structure localisation of single vertebrae in CT scans for a subsequent model initialisation of more sophisticated segmentation algorithms. A MRF based model of appearance, which employs local information in regions around anatomical landmarks and geometrical information through connections between adjacent landmarks, is built on volumetric CT datasets of lumbar vertebrae. The MRF based model is built on a 6 landmark configuration in vertebra volumetric data and is additionally matched with target data. This is done by finding a best fit MRF matching by the Max-sum algorithm among feature points found by a decision tree based feature detection algorithm called probabilistic boosting tree (PBT). Anatomical landmark regions are described by vector spin-images and shape index histograms. Adjacency information is extracted by Delaunay tetrahedralisation where distances and gradient-related angles describe connections between adjacent regions. The results on single lumbar vertebra CT scans show that the MRF approach is applicable on volumetric CT datasets with an accuracy enough for supporting more sophisticated segmentation algorithms such as Active Appearance Models (AAMs).

Additional Files and Images

Additional images and videos

Additional files

Weblinks

No further information available.

BibTeX

@mastersthesis{major-2010-mrf,
  title =      "Markov Random Field Based Structure Localisation of
               Vertebrae for 3D-Segmentation of the Spine in CT Volume Data",
  author =     "David Major",
  year =       "2010",
  abstract =   "Medical Image Processing is a growing field in medicine and
               plays an important role in medical decision making.
               Computer-based segmentation of anatomies in data made by
               imaging modalities supports clinicians and speeds up their
               diagnosis making compared to doing it manually. Computed
               Tomography (CT) is an imaging modality for slice-wise three
               dimensional reconstruction of the human body in the form of
               volumetric data which is especially applicable for imaging
               of bony structures and so for the vertebral column. Most
               bony structures, such as vertebrae, are characterised by
               complex shape and texture appearances which turns its
               segmentation into a difficult task. Model-based segmentation
               approaches are promising techniques to cope with variations
               in form and texture of the anatomy of interest. This is done
               by incorporating information about shape and texture
               appearance gained from an imaging modality in a model. The
               model can then be applied to segment the object of interest
               in target data, however most of the model-based approaches
               need a model intialisation for a fast and reliable
               segmentation of the object of interest. This thesis was
               motivated by novel works on fast anatomical structure
               localisation with Markov Random Fields (MRFs) and focuses on
               the sparse structure localisation of single vertebrae in CT
               scans for a subsequent model initialisation of more
               sophisticated segmentation algorithms. A MRF based model of
               appearance, which employs local information in regions
               around anatomical landmarks and geometrical information
               through connections between adjacent landmarks, is built on
               volumetric CT datasets of lumbar vertebrae. The MRF based
               model is built on a 6 landmark configuration in vertebra
               volumetric data and is additionally matched with target
               data. This is done by finding a best fit MRF matching by the
               Max-sum algorithm among feature points found by a decision
               tree based feature detection algorithm called probabilistic
               boosting tree (PBT). Anatomical landmark regions are
               described by vector spin-images and shape index histograms.
               Adjacency information is extracted by Delaunay
               tetrahedralisation where distances and gradient-related
               angles describe connections between adjacent regions. The
               results on single lumbar vertebra CT scans show that the MRF
               approach is applicable on volumetric CT datasets with an
               accuracy enough for supporting more sophisticated
               segmentation algorithms such as Active Appearance Models
               (AAMs).",
  month =      may,
  address =    "Favoritenstrasse 9-11/186, A-1040 Vienna, Austria",
  school =     "Institute of Computer Graphics and Algorithms, Vienna
               University of Technology",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2010/major-2010-mrf/",
}