Information

  • Publication Type: Master Thesis
  • Workgroup(s)/Project(s): not specified
  • Date: October 2022
  • Date (Start): 2021
  • Date (End): 2022
  • Diploma Examination: 10. October 2022
  • First Supervisor: Renata RaidouORCID iD
  • Pages: 148
  • Keywords: Convolutional Neural Network, Prediction, Knee Osteoarthritis

Abstract

Osteoarthritis (OA) is a slowly degenerative joint disease, with cartilage loss as one of the most characteristic symptoms accompanied by pain and functional disability. The knee region is the most affected area. 22.9% of the worldwide population over the age of 40 were affected in 2020 by Knee Osteoarthritis (KOA). Besides normal KOA, which develops over multiple years, the accelerated form of KOA (AKOA) develops between 1 and 4 years and is accompanied by increased pain and movement restrictions as well as a higher chance of obtaining a knee replacement. The development of AKOA is not yet predictable on the basis of a single X-ray image because there is no apparent optical difference between the baseline X-ray of KOA and AKOA. Since Convolutional Neural Networks (CNN) can identify image structures that a human eye can not see, I want to realise an early diagnosis of AKOA by using a Convolutional Neural Network (CNN) as a classifier between slow- and fast-progressing KOA.For this purpose, I used the data from three different studies, including a knee X-ray, Body Mass Index (BMI), age, gender, Western Ontario and McMaster Universities Arthritis Index (WOMAC) scores, hip symptoms, knee medication injection and Kellgren- Lawrence (KL)-grade, as input for binary classification models. I defined AKOA once with Joint Space Narrowing (JSN) > 10%/ 2 years and once with JSN > 20%/ 2 years and performed different experiments in order to find the best method to predict AKOA. I trained the numeric data only on an Extreme Gradient Boosting (XGBoost) model. Here I achieved the highest performance of an Area Under the Curve (AUC) of 0.6616 when including the Osteoarthritis Research Society International (OARSI) score of sclerosis and osteophytosis to the numeric input data (20% JSN/ 2 years). To use image data only and the combination of both, I created different CNN models, whose architecture is based on a Residual Network (ResNet) 50 model provided by Image Biopsy Lab (IBLab). The CNN model, which I trained only with image data, yielded an AUC of 56.26% (10% JSN/ 2 years). Using the image data complemented with the most important numeric features (gender, BMI, contralateral KOA, KL-grade) as input, I achieved an AUC of 68.78% (20% JSN/ 2 years). Comparable results, but obtained with other class definitions than in this work, were higher and yielded AUCs of around 0.8.These results show that it is possible to make a risk assessment about the development of AKOA using the baseline X-ray image, gender, BMI, the KL-grade and the information about contralateral KOA. Until now, radiologists are not capable of predicting fast-progressing KOA. Hence, these networks have a great potential to be used as AKOA prediction tools.

Additional Files and Images

Additional images and videos


Additional files

Weblinks

BibTeX

@mastersthesis{vogel2022,
  title =      "Prediction of Accelerated Knee Osteoarthritis Using a
               Convolutional Neural Network",
  author =     "Magdalena Vogel",
  year =       "2022",
  abstract =   "Osteoarthritis (OA) is a slowly degenerative joint disease,
               with cartilage loss as one of the most characteristic
               symptoms accompanied by pain and functional disability. The
               knee region is the most affected area. 22.9% of the
               worldwide population over the age of 40 were affected in
               2020 by Knee Osteoarthritis (KOA). Besides normal KOA, which
               develops over multiple years, the accelerated form of KOA
               (AKOA) develops between 1 and 4 years and is accompanied by
               increased pain and movement restrictions as well as a higher
               chance of obtaining a knee replacement. The development of
               AKOA is not yet predictable on the basis of a single X-ray
               image because there is no apparent optical difference
               between the baseline X-ray of KOA and AKOA. Since
               Convolutional Neural Networks (CNN) can identify image
               structures that a human eye can not see, I want to realise
               an early diagnosis of AKOA by using a Convolutional Neural
               Network (CNN) as a classifier between slow- and
               fast-progressing KOA.For this purpose, I used the data from
               three different studies, including a knee X-ray, Body Mass
               Index (BMI), age, gender, Western Ontario and McMaster
               Universities Arthritis Index (WOMAC) scores, hip symptoms,
               knee medication injection and Kellgren- Lawrence (KL)-grade,
               as input for binary classification models. I defined AKOA
               once with Joint Space Narrowing (JSN) > 10%/ 2 years and
               once with JSN > 20%/ 2 years and performed different
               experiments in order to find the best method to predict
               AKOA. I trained the numeric data only on an Extreme Gradient
               Boosting (XGBoost) model. Here I achieved the highest
               performance of an Area Under the Curve (AUC) of 0.6616 when
               including the Osteoarthritis Research Society International
               (OARSI) score of sclerosis and osteophytosis to the numeric
               input data (20% JSN/ 2 years). To use image data only and
               the combination of both, I created different CNN models,
               whose architecture is based on a Residual Network (ResNet)
               50 model provided by Image Biopsy Lab (IBLab). The CNN
               model, which I trained only with image data, yielded an AUC
               of 56.26% (10% JSN/ 2 years). Using the image data
               complemented with the most important numeric features
               (gender, BMI, contralateral KOA, KL-grade) as input, I
               achieved an AUC of 68.78% (20% JSN/ 2 years). Comparable
               results, but obtained with other class definitions than in
               this work, were higher and yielded AUCs of around 0.8.These
               results show that it is possible to make a risk assessment
               about the development of AKOA using the baseline X-ray
               image, gender, BMI, the KL-grade and the information about
               contralateral KOA. Until now, radiologists are not capable
               of predicting fast-progressing KOA. Hence, these networks
               have a great potential to be used as AKOA prediction tools.",
  month =      oct,
  pages =      "148",
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien",
  keywords =   "Convolutional Neural Network, Prediction, Knee
               Osteoarthritis",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2022/vogel2022/",
}