PhD Defence: COMPUTED TOMOGRAPHY TEXTURE-BASED RADIOMICS WITH MACHINE LEARNING FOR PREDICTING CHRONIC OBSTRUCTIVE PULMONARY DISEASE OUTCOMES

Student: Kalysta Makimoto

Supervisor: Dr. Miranda Kirby

Abstract

Chronic obstructive pulmonary disease (COPD) is characterized by the presence of emphysema and/or chronic bronchitis, which can be quantified using computed tomography (CT) imaging. Conventional CT imaging features used for COPD provide global measurements and do not provide information about the distribution of the disease. Texture-based radiomics overcomes this limitation by capturing the spatial distribution of the grey levels in the CT images. However, prior to the work in this dissertation, CT texture-based radiomics were not widely investigated for predicting COPD outcomes with machine learning (ML) models.

The aim of Chapter 2 was to develop a ML pipeline for predicting COPD status using texture- based radiomics. An optimal ML pipeline was identified, including an Elastic Net feature selection technique and Support Vector Machine classifier. Additionally, the ML model utilizing the texture- based radiomics obtained improved performance compared to a traditional model with conventional CT imaging features.

The aim of Chapter 3 was to implement the developed ML pipeline and investigate if the combination of texture-based radiomics and conventional CT imaging features obtained improved prediction performance compared to either feature set alone. These results demonstrated that the addition of texture-based radiomics to conventional CT imaging features obtained improved performance for prediction COPD status and severity.

The aim of Chapter 4 was to predict incident COPD in at-risk individuals using the developed ML pipeline. The ML model utilizing clinical features and CT imaging features, texture-based radiomics and conventional CT imaging features, obtained improved prediction performance compared to existing models.

The aim of Chapter 5 was to predict lung function decline using the developed ML pipeline and upper-lobe CT imaging features. These results demonstrated that in at-risk individuals, upper-lobe CT imaging features were more predictive of lung function decline compared to whole lung features.

The aim of Chapter 6 was to develop sex-specific ML models for predicting incident and prevalent COPD. The female cohort model obtained significantly improved performance for predicting incident and prevalent COPD, as well as identified sex-specific features. Overall, the projects in this dissertation developed ML models that utilized CT texture-based radiomics and demonstrated their ability and importance for predicting COPD outcomes.