Information

Abstract

This master thesis aims to provide an in-depth comparison of four texture algorithms in their capacity of discriminating patients with osteoarthritis (OA) from the ones without, recognizing early signs of Osteoarthritis and tracking disease progression from 2D radiographs of the knee trabecular bone (TB). Given the fractal properties of the trabecular bone (TB), two fractal-based algorithms (Bone Variance Value (BVV) and Bone Score Value (BSV)) that try to characterize the complexity of the underlying 3D structure of the bone are presented. The third algorithm (Bone Entropy Value (BEV), based on Shannon’s Entropy) stems from the information theory and aims to describe the bone structure in terms of information complexity. The last algorithm (Bone Coocurrence Value (BCV)) is based on the co-occurrence matrix of an image and describes the image texture in terms of certain Haralick features. If successful, such algorithms posses a great potential to lower the costs (financial, time) associated with the diagnosis of osteoarthritis (OA) through automation of the procedure, and with the treatment. The earlier treatments and risk reduction measures are less costly than the procedures involved due to a more advanced stage of the disease (surgery, implants, etc.). First, a motivation for the detection of early osteoarthritis (OA) is given. Second, a detailed description and mathematical background of the algorithms are presented and validated on sample, artificial data. Third, the employed data sets used for classification tests are introduced. Fourth, the statistical methods and neural network models employed are presented and discussed. Fifth, the features produced by each algorithm are discussed and their independent and combined capacity of discriminating between bones with early signs of OA and healthy bones. Also the capacity of tracking OA progression through the years is quantified by statistical tests. Also in this part we present the best classification scores obtained from the most optimal neural networks for each use case. Finally, thoughts on future improvements and the generalization of the algorithms in other anatomical contexts, for other diseases or in other fields, like histology and mammography, are made. In this work we show that the state-of-the-art in OA prediction can be surpassed by utilizing only models based on texture features alone. Our gender-stratified analysis produces a prediction score of 83% for males and 81% for females in terms of Area Under the Receiver Operating Characteristic Curve (ROC-AUC).

Additional Files and Images

Additional images and videos

Additional files

Weblinks

No further information available.

BibTeX

@mastersthesis{Oancea_2018_1,
  title =      "Four Texture Algorithms for Recognizing Early Signs of
               Osteoarthritis. Data from the Multicenter Osteoarthritis
               Study.",
  author =     "Stefan Ovidiu Oancea",
  year =       "2018",
  abstract =   "This master thesis aims to provide an in-depth comparison of
               four texture algorithms in their capacity of discriminating
               patients with osteoarthritis (OA) from the ones without,
               recognizing early signs of Osteoarthritis and tracking
               disease progression from 2D radiographs of the knee
               trabecular bone (TB). Given the fractal properties of the
               trabecular bone (TB), two fractal-based algorithms (Bone
               Variance Value (BVV) and Bone Score Value (BSV)) that try to
               characterize the complexity of the underlying 3D structure
               of the bone are presented. The third algorithm (Bone Entropy
               Value (BEV), based on Shannon’s Entropy) stems from the
               information theory and aims to describe the bone structure
               in terms of information complexity. The last algorithm (Bone
               Coocurrence Value (BCV)) is based on the co-occurrence
               matrix of an image and describes the image texture in terms
               of certain Haralick features. If successful, such algorithms
               posses a great potential to lower the costs (financial,
               time) associated with the diagnosis of osteoarthritis (OA)
               through automation of the procedure, and with the treatment.
               The earlier treatments and risk reduction measures are less
               costly than the procedures involved due to a more advanced
               stage of the disease (surgery, implants, etc.). First, a
               motivation for the detection of early osteoarthritis (OA) is
               given. Second, a detailed description and mathematical
               background of the algorithms are presented and validated on
               sample, artificial data. Third, the employed data sets used
               for classification tests are introduced. Fourth, the
               statistical methods and neural network models employed are
               presented and discussed. Fifth, the features produced by
               each algorithm are discussed and their independent and
               combined capacity of discriminating between bones with early
               signs of OA and healthy bones. Also the capacity of tracking
               OA progression through the years is quantified by
               statistical tests. Also in this part we present the best
               classification scores obtained from the most optimal neural
               networks for each use case. Finally, thoughts on future
               improvements and the generalization of the algorithms in
               other anatomical contexts, for other diseases or in other
               fields, like histology and mammography, are made. In this
               work we show that the state-of-the-art in OA prediction can
               be surpassed by utilizing only models based on texture
               features alone. Our gender-stratified analysis produces a
               prediction score of 83% for males and 81% for females in
               terms of Area Under the Receiver Operating Characteristic
               Curve (ROC-AUC).",
  month =      jun,
  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/2018/Oancea_2018_1/",
}