Exploring and understanding the impact of machine learning choices on radiotherapy decision making

Information

• Publication Type: Master Thesis
• Workgroup(s)/Project(s): not specified
• Date: 2023
• Open Access: yes
• First Supervisor: Renata Raidou
• Pages: 107
• Keywords: machine learning, prediction, statistical evaluation, clustering, shape descriptor, CT scan, visual analytics, prostate cancer

Abstract

In prostate cancer radiotherapy planning, the accurate description of the position and shape of pelvic organs is a crucial part of successful patient treatment. However, the treatment is conducted throughout a long period of time, during which the position and shape of the organs might significantly vary. In addition, the amount of variation tends to differ for each individual. Recent visual analytics publications investigated this by partitioning past patients into clusters with similar variability. Using this as part of a prediction for the organ variability of new patients could improve and further personalize therapy planning. However, the statistical and machine learning methods employed in these works have not been thoroughly and quantitatively evaluated so far and their impact on the final predictions has not been assessed. This thesis focuses on taking a particular implementation of these approaches, proposed by Furmanová et al. [FMCM+21], and quantitatively evaluating the effects of using different alternatives for the employed methods. We focus on two aspects: the effect of using different shape descriptor methods and the impact of modifications in the clustering methods employed. By providing an additional visual analytics framework to visually assess the effect of the aforementioned alternatives, we aim to ensure an effortless and interactive visual interpretation of the impact of various modifications. This is anticipated to support the developers of said predictive algorithms in designing more robust approaches. As a result of our investigation we have highlighted potential issues and improved the initial implementation of the proposed workflow. We conclude that at the current stage of the patient cohort used for the analysis, the selection of appropriate shape description methods should be of main focus, while a notable impact of using different clustering methods is limited to the prediction of the most extreme cases of organ shape variations.

Additional Files and Images

Weblinks

• Entry in reposiTUm (TU Wien Publication Database)
• DOI: 10.34726/hss.2023.91282

BibTeX

@mastersthesis{boeroendy-2023-eui,
  title =      "Exploring and understanding the impact of machine learning
               choices on radiotherapy decision making",
  author =     "Ádam B\"{o}r\"{o}ndy",
  year =       "2023",
  abstract =   "In prostate cancer radiotherapy planning, the accurate
               description of the position and shape of pelvic organs is a
               crucial part of successful patient treatment. However, the
               treatment is conducted throughout a long period of time,
               during which the position and shape of the organs might
               significantly vary. In addition, the amount of variation
               tends to differ for each individual. Recent visual analytics
               publications investigated this by partitioning past patients
               into clusters with similar variability. Using this as part
               of a prediction for the organ variability of new patients
               could improve and further personalize therapy planning.
               However, the statistical and machine learning methods
               employed in these works have not been thoroughly and
               quantitatively evaluated so far and their impact on the
               final predictions has not been assessed. This thesis focuses
               on taking a particular implementation of these approaches,
               proposed by Furmanov\'{a} et al. [FMCM+21], and
               quantitatively evaluating the effects of using different
               alternatives for the employed methods. We focus on two
               aspects: the effect of using different shape descriptor
               methods and the impact of modifications in the clustering
               methods employed. By providing an additional visual
               analytics framework to visually assess the effect of the
               aforementioned alternatives, we aim to ensure an effortless
               and interactive visual interpretation of the impact of
               various modifications. This is anticipated to support the
               developers of said predictive algorithms in designing more
               robust approaches. As a result of our investigation we have
               highlighted potential issues and improved the initial
               implementation of the proposed workflow. We conclude that at
               the current stage of the patient cohort used for the
               analysis, the selection of appropriate shape description
               methods should be of main focus, while a notable impact of
               using different clustering methods is limited to the
               prediction of the most extreme cases of organ shape
               variations.",
  pages =      "107",
  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 =   "machine learning, prediction, statistical evaluation,
               clustering, shape descriptor, CT scan, visual analytics,
               prostate cancer",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2023/boeroendy-2023-eui/",
}