Heterogeneous radiotherapy dose-outcomes response in parotid glands

Abstract

Introduction: Parotid glands are treated clinically as though the distribution of functional burden were homogeneous. Radiotherapy treatments are planned using whole parotid mean dose to predict risk of salivary dysfunction. Recent progress has identified specific parotid non-homogeneities by demonstrating the existence of regional, bath-and-shower, and dose-volume effects. In this work, parotid regional effects and their impact on salivary function are quantified using a non-parametric (model-free) approach. Regional effects have implications for clinical sparing practices. Materials and Methods: Radiotherapy planning contours, dose profiles, and late clinical outcomes from a single cohort consisting of N = 332 patients was used. Pre-radiotherapy and one year post-radiotherapy whole mouth stimulated saliva were collected for assessment of salivary dysfunction. Organ-at-risk parotid glands were segmented into 2, 3, 4, 18, and 96 equal-volume sub-segments. Sub-segment relative importance was derived from mean dose regressors using random forests and conditional inference trees. Regressor multicollinearity, cohort homogeneity, and overfitting were addressed. Linear and exponential whole parotid mean dose models were also implemented for comparison purposes. Results: Exclusion of caudal-anterior sub-segments negatively impacted prediction the most. The most important sub-segments had importances 2.4× (on average over all segmentation methods) or >4× (at the finest level of segmentation) that of an equivalent sub-segment in a theoretical homogeneous parotid. In contrast, the least important sub-segments held virtually no importance for prediction. Both random forests and conditional inference trees outperformed parametric (model-based) techniques. Both improved prediction as segmentation was refined. Conclusions: Radiation dose to caudal-anterior aspects of the parotid are the strongest predictors of radiotherapy-induced late stimulated whole mouth saliva, and are thus the most clinically-relevant regions for controlling dysfunction. Cranial and posterior aspects are less important. Shifting dose from regions of high importance to low importance may therefore improve patient outcomes.