Radiomics-based machine learning for predicting non-small cell lung cancer outcomes

Radiomics is a quantitative medical image analysis method that can be used to extract underlying patterns in computed tomography (CT) scans. Radiomic features may capture information about cancer biology and may be used to improve cancer prognosis and predict treatment response. This thesis aims to leverage radiomics and machine learning (ML) to predict non-small cell lung cancer (NSCLC) patient outcomes, ultimately working towards supporting personalized medicine. In this work, radiomic feature consistency across provincial institutions was evaluated using a novel radiomic phantom. Radiomic features are sensitive to imaging parameters, so this crucial first step ensures the robustness of downstream analysis involving multi-institutional data. The most stable features were identified for outcomes modelling. Next, explainable artificial intelligence (XAI) models to predict NSCLC recurrence based on radiomic and clinical features were developed. Although ML models can predict outcomes, the reasoning behind their predictions is often unclear. This preliminary study identified the most informative features for predicting recurrence, demonstrating the utility of XAI for understanding model decisions. Then, ML models to predict post-stereotactic ablative radiotherapy (SABR) NSCLC recurrence were trained on different feature types. These included radiomic features and clinical information, as well as dosimetric features from radiotherapy plans. In addition, so-called foundation features identified by a largescale ML system were considered. The predictive power of multi-modal features was compared. Finally, an approach for outcomes modelling using a synthetic dataset was proposed to address the need for greater data availability. A comprehensive synthetic dataset containing radiomic and clinical features was generated and then used to develop ML models to predict NSCLC survival. Although shown in the context of NSCLC survival, the methodology could be applied to other use cases.