EP-2121: Serial Parotid Gland Radiomic-based Model Predicts Post-Radiation Xerostomia in Oropharyngeal Cancer

Abstract

Purpose or Objective The aim of this work is to investigate the applicability of combining serially derived radiomic features with standard clinical variables into a xerostomia probability computational tool. Material and Methods Patients undergoing RT for oropharyngeal cancer with daily CT-on-rails imaging were reviewed. Xerostomia status at 6-months post-treatment was retrieved as well as chief clinical variables, (e.g. age, smoking status, Tcategory, chemotherapy, and dose received). A total of 437 scans were analyzed. Ipsilateral parotid glands were contoured on baseline CT and propagated to daily CT images using deformable vector fields generated for IGRT until mid-treatment time point. A total of 145 radiomic features were selected from the categories: Intensity direct, neighborhood intensity difference, grey-level cooccurrence matrix, grey-level run length and shape. Spearman correlation was used to reduce the 145 features to 5 features based on a cutoff of 0.7. These features included: 'LocalEntropyStd', 'LocalStdStd', 'Compactness2', 'Volume' and 'Contrast'. These were then used to build and evaluate three distinct types of models (1) using only the baseline (BL) value of the radiomic feature, (2) the ratio between the mid-RT value of the radiomics feature to its initial value at BL, and (3) a functional principal component analysis (FPCA) model that leverages the structure of the temporal trajectory in the evolution of the radiomic features from BL to mid-RT. Afterwards, we ran logistic regression was to explore the capacity of relevant clinical variables to predict xerostomia at 6-month post-RT; either alone or in combination with one of the radiomic-derived models. Results 28 patients were included. At 6 months, xerostomia was reported as follows: Minimal or mild (70.4%) vs moderate-severe (29.6%). The corresponding Receiver operating characteristics area under the curve (ROC AUCs) and confidence intervals (CIs) for various predictive models plotted and depicted in Figure & Table 1. Clinical only model showed worse AUC when compared with any composite clinical/radiomics models. Combination radiomics model included input from all 3 radiomics models: Baseline, delta & FPCA. Noteworthy, each of the 3 individual radiomics models performed worse than the clinical model with corresponding AUCs of: 0.59, 0.64, and 0.72, respectively. This suggests that the functional approach yields a superior ROC compared to using either the baseline radiomics feature or using the delta ratio between the mid and initial time points. Conclusion Textural kinetic trajectories from consequential intratreatment CT scans can predict for subsequent radiationinduced toxicities. Combining clinical and radiomics input can synergistically add up to the predictive capacity of post-RT xerostomia probability computation. (Figure Presented).