Semi-Automatic Spine Labeling on T1- and T2-weighted MRI Volume Data | TU Wien

Information

Publication Type: Master Thesis
Workgroup(s)/Project(s):
- VRVis Competence Center
Date: January 2015
Second Supervisor: Katja Bühler
First Supervisor:
- Eduard Gröller
- Katja Bühler

Abstract

In medical diagnosis, the spine is often a frame of reference and so helps to localize diseases (e.g. tumors) in the human body. Automated spine labeling approaches are in demand, in order to replace time consuming, manual labeling by a radiologist. Different approaches have already been proposed in the literature, mainly for Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) data. While CT scans exhibit a generalized intensity scale, MR images come with a high variability within the data and hence the tissues. Several factors influence the appearance of vertebrae and intervertebral disks in MRI data: different scanners, changes of acquisition parameters, magnetic field inhomogeneities or age-related, structural changes of the spinal anatomy. These factors compound the development of semi- and fully automatic spine labeling systems.

The main goal of this thesis is to overcome these variations and find a generalized representation for different kinds of MR data. Furthermore, it aims for a semi-automatic labeling approach on these preprocessed scans where the user has to provide an initial click. Entropyoptimized Texture Models are applied to normalize the data to a standardized, reduced intensity scale.With Probabilistic Boosting Trees, intervertebral disk feature points are detected, whereby the disk center is selected with a Shape Particle Filter.

The results achieved with the proposed pipeline are promising in terms of data normalization, timing and labeling accuracy. With a mean overall processing time of 6.0 s for normalizing and labeling a dataset (0.8 s per disk), the algorithm achieves a precision of 92.4% (recall = 86.8%). Using a higher resolution of the data for disk detection (average timing of 1.6 s per disk resp. 12.4 s per dataset), reduces the number of missed disk candidates and hence increases the recall to 91.7% (with a precision of 91.9%).

Additional Files and Images

Poster

Thesis

Weblinks

No further information available.

BibTeX

@mastersthesis{Wimmer_Maria_2015_SAS,
  title =      "Semi-Automatic Spine Labeling on T1- and T2-weighted MRI
               Volume Data",
  author =     "Maria Wimmer",
  year =       "2015",
  abstract =   "In medical diagnosis, the spine is often a frame of
               reference and so helps to localize diseases (e.g. tumors) in
               the human body. Automated spine labeling approaches are in
               demand, in order to replace time consuming, manual labeling
               by a radiologist. Different approaches have already been
               proposed in the literature, mainly for Computed Tomography
               (CT) and Magnetic Resonance Imaging (MRI) data. While CT
               scans exhibit a generalized intensity scale, MR images come
               with a high variability within the data and hence the
               tissues. Several factors influence the appearance of
               vertebrae and intervertebral disks in MRI data: different
               scanners, changes of acquisition parameters, magnetic field
               inhomogeneities or age-related, structural changes of the
               spinal anatomy. These factors compound the development of
               semi- and fully automatic spine labeling systems.  The main
               goal of this thesis is to overcome these variations and find
               a generalized representation for different kinds of MR data.
               Furthermore, it aims for a semi-automatic labeling approach
               on these preprocessed scans where the user has to provide an
               initial click. Entropyoptimized Texture Models are applied
               to normalize the data to a standardized, reduced intensity
               scale.With Probabilistic Boosting Trees, intervertebral disk
               feature points are detected, whereby the disk center is
               selected with a Shape Particle Filter.  The results achieved
               with the proposed pipeline are promising in terms of data
               normalization, timing and labeling accuracy. With a mean
               overall processing time of 6.0 s for normalizing and
               labeling a dataset (0.8 s per disk), the algorithm achieves
               a precision of 92.4% (recall = 86.8%).  Using a higher
               resolution of the data for disk detection (average timing of
               1.6 s per disk resp. 12.4 s per dataset), reduces the number
               of missed disk candidates and hence increases the recall to
               91.7% (with a precision of 91.9%).",
  month =      jan,
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Institute of Computer Graphics and Algorithms, Vienna
               University of Technology ",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2015/Wimmer_Maria_2015_SAS/",
}