Towards Robust Radiomic Markers from Positron Emission Tomography in Cancer

Abstract

[18F]-fluorodeoxyglucose positron emission tomography ([18F]-FDG PET) images contain both rich and untapped information regarding diseases such as cancer. The latter can be accessed computationally using radiomics, from which we can build new markers of disease to improve the clinical management of patients. However, the generation and handling of PET images is a multifactorial process, and so careful consideration of these factors is required to establish reliable radiomic markers. This PhD thesis begins with an overview of the theory behind PET imaging in cancer and radiomics. It then provides a systematic review of works that have preceded this thesis, where the methodological characteristics and limitations of existing robustness investigations involving PET radiomics are highlighted. Informed by this review, three robustness studies are subsequently presented, which collectively investigate the response of [18F]-FDG PET radiomic features to various imaging scenarios. The first study provides an in-depth evaluation of how variations in intensity discretisation and voxel interpolation parameters affect radiomic features in non-small cell lung cancer, melanoma, and lymphoma. There, two strategies, i.e., implementing parameter range conditions and dependency corrections, are introduced to potentially mitigate the instability of feature values. The effectiveness of the latter is further explored in a second study, which examines [18F]-FDG PET radiomic features obtained from an anthropomorphic phantom and their robustness to different acquisition and reconstruction settings. The final study evaluates the impact of harmonisation using ComBat, an approach based on empirical Bayes, on the robustness of [18F]-FDG PET radiomic features across discordant imaging parameters. It focuses on six generalised variants of the ComBat method and their influence on predictive markers of overall survival in head-and-neck cancer. This work builds on existing literature by providing a more nuanced understanding of the behaviour of [18F]-FDG PET radiomics data in response to different imaging parameters, as well as to strategies that could mitigate feature non-robustness. It also reinforces the need for mindful selection of PET imaging markers prior to their deployment in high-stakes situations like cancer healthcare.