Quantitative Image Analysis for the Assessment of Parotid Gland Damage and Toxicity Prediction Following Radiotherapy to the Head and Neck

Abstract

Patients treated with radiotherapy for head and neck cancer experience debilitating xerostomia and sticky saliva, caused by damage to the parotid glands. Currently predicting this toxicity is limited by the inability of existing normal tissue complication probability models to account for the complexity of the glands’ patient specific anatomy and their heterogeneous biology. Applying quantitative imaging analysis via “radiomics” enables medical images to be converted to minable data which reflects underlying biology of specific regions of interest. These radiomic features could be informative of how normal tissues respond to treatment by documenting patient specific texture of the parotid glands. Hypothetically, radiation-induced xerostomia and sticky saliva could be predicted prior to the manifestation of clinical signs, using radiomic analysis of routine computer tomography (CT) images from patients with head and neck cancer. Prior to testing this hypothesis, this thesis first investigated specific nuances within the radiomics pipeline which could influence radiomic feature values. Specifically, Study 1 quantified the difference in radiomic feature values that can occur as a result of inter-observer delineation variability of the right parotid gland. Based on a cohort of 40 independent observers, results of this study identified 14 specific radiomic features which are robust to these uncertainties. It is understood that radiomic features are also influenced by CT scanner acquisition and reconstruction protocols. This variability can be assuaged by pre-processing methods applied prior to feature extraction, for example via spatial resampling and intensity discretisation; however, a knowledge gap existed for this approach when studying the parotid glands. Study 2 therefore explored the impact of resampling to a common voxel size on radiomic feature values. This process was tested across two different bin widths to assess the reproducibility under different discretisation conditions. Based on a cohort of 127 patients, results of this study, identified several radiomic features which are robust to these pre-processing settings. Results here indicated the combination of isotropic resampling and the use of a smaller bin width of 5 HU may be optimal for radiomic analysis of the parotid glands. The work conducted in Studies 1 and 2 was a required step for the responsible exploration of the radiomics pipeline; however, they lacked an application to clinical outcomes. This clinical link was subsequently addressed in Studies 3 and 4. In Study 3 logistic regression models were built using a combination of clinical, dosimetric, and radiomic features to predict acute xerostomia and sticky saliva. This study included a cohort of 98 patients treated with chemoradiation for squamous cell carcinoma of the head and neck. The integration of radiomics into the models improved the prediction of acute xerostomia and sticky saliva. Radiation dose is accumulated during a course of treatment, and the subsequent tissue response is therefore dynamic. Based on images acquired at three timepoints in a cohort of 26 patients, Study 4 identified the variability of specific radiomic features that were associated with acute and late xerostomia. Quantifying variation in radiomic features prior to the onset of clinical symptoms in such a way, allows the opportunity for prompt intervention, which may include adaptive radiotherapy. The work presented in this thesis has demonstrated the viability of radiomic analysis of parotid gland damage following radiotherapy for head and neck cancer. The outcome of this work will ultimately assist in the personalisation of radiotherapy and supportive care for our patients.