| dc.description.abstract | Radiation therapy (RT) is often included in, or used as a stand-alone, treatment of cancer in the head and neck region (HNC). The salivary glands (SGs) are often in close proximity to the tumour and are therefore not always possible to spare from irradiation when using RT to treat HNC. The response of the SGs to irradiation show variations between patients, indicating that some are more radiosensitive than others. Incorporating more patient-specific biomarkers into treatment planning and evaluations during fractionated radiotherapy are needed to establish a precision oncology framework for mitigation of side-effects such as xerostomia (dry mouth). Radiomic image features from medical imaging have previously shown potential as biomarkers for risk of developing xerostomia post-RT. Radiomics is a high-throughput method of extracting quantitative information from such images and the features may be broadly categorized as shape-based, first-order, and texture-based. This work evaluated the relation between 828 radiomic features calculated from 2D regions of interests (ROIs) in either T1- or T2-weighted magnetic resonance images (MRI) to known biological changes in the SGs due to damage by ionizing radiation. C57BL/6J mice were used, where 72 individuals were irradiated while 40 belonged to a control group. The developed radiomic workflow includes creation of ROIs for each image (segmentation), preprocessing, feature extraction, feature selection, and modelling. Radiomic studies are known to have an issue with reproducibility and feature robustness, and therefore all steps in the workflow are evaluated and described in detail. Intensity normalization was performed on a feature-specific level. The segmented ROIs were evaluated against sublingual gland (SLG) and submandibular gland (SMG) areas from 9 surgical specimens. While the SLG areas had higher correlation to the image-segmented ROIs than the SMG areas (ρ=0.71 and 0.36, respectively), the two types of glands could not be differentiated in the images due to being fused in mice. Saliva production was found to be significantly lower in irradiated individuals relative control comparing data from between day 26 and 105 post-irradiation. Xerostomia was defined into a binary outcome variable by thresholding. Using only image features from T1 images proved to be significantly better predictors of xerostomia than the T2 features when evaluated on the same data. The relative difference in a T2 first-order feature before and after pilocarpine injections for saliva measurements, delta-p energy, was shown to be a high-performing predictor of xerostomia evaluated on the same day as the MR imaging. Overall, 2D features from the right SG-subunit proved to be higher-performing features than the left subunit, possibly due to some differences in delivered dose. Both T1 and T2 features obtained from MRI after irradiation were good predictors of late xerostomia, but only T1 features showed a possible predictive ability at baseline. The relative difference in a shape-feature before and after irradiation (delta-feature) showed promise of predicting late xerostomia. Multiple textural features from both T1 and T2 images were good predictors of late xerostomia, possibly related to changes in vascularity or increased fatty tissue in the glands post-irradiation. The radiomic image features were able to predict saliva production in C57BL/6J mice with varying accuracy. 14 features significantly improved upon models only using time and dose as predictors, indicating that certain features contained information relating to the inter-mouse variations affecting saliva production. However, none of the features were significant under Bonferroni corrected p-threshold, emphasizing the need for validating studies on external data. | eng |