Thesis
Image processing and machine learning applications in lung cancer treatment
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17257
- Person Identifier (Local)
- 202063066
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- The continued advancement of image processing and machine learning techniques opens up the opportunity for their application in the medical setting. The aim of the work in this thesis was to apply these techniques from this broad field to lung cancer treatment with the aim of providing tools that can improve patient outcomes. The topics covered were; pulmonary and esophageal toxicity following radiotherapy, registration of PET/CT imaging to pathology and the automatic segmentation of tumour regions in gross pathology images. For the prediction of pulmonary toxicity, predictive features were extracted from pre-treatment planning CT images using radiomic and deep learning based approaches. When combined with dose features, these models produced a large increase in predictive power compared to models using only dose and clinical features. For the ILD patients receiving SABR, predictive power was also shown on several metrics such as the FACT-L and EQ-5D-5L scales. For predicting esophageal toxicity, the data from the RTOG0617 clinical trial was used. Here the focus was on improving predictions from the dose maps. It was found that using 3D-CNNs, regression based training, including additional toxicities and ensembling models improved model performance. Tests were also conducted to determine the robustness of boosted decision tree and artificial neural network based models for esophageal toxicity prediction by adding noise to the test data. The PET/CT to pathology registration task followed on from a previous project that built the framework for registering CT to pathology but failed to include PET due to respiratory motion blurring. This was added to by including respiratory gating and the OncoFreeze algorithm in the workflow to reduce the effects of respiratory motion. A PET to pathology registration was evaluated using thresholding based registration of the PET image. Additionally, a deep learning based method for the automatic segmentation of gross pathology images was produced. This included training and testing various UNet and DeeplabV3+ models with both Dice and cross entropy based loss functions. The best performing model was an ensemble of several models with morphological post processing steps.
- Advisor / supervisor
- Harrow, Stephen
- Marshall, Stephen
- Murray, Paul
- Dick, Craig
- Resource Type
- DOI
- Date Created
- 2024
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T17257 | 2025-03-11 | Public | Download |