Role of Imaging in Predicting Radiation Induced Xerostomia

Abstract

The development of predictive biomarkers for radiation toxicities (e.g. xerostomia) may better identify head and neck cancer (HNC) patients suitable for treatment intensification. Radiomics is a promising approach to identify image-based biomarkers. We hypothesize that baseline salivary gland structure-function evaluated using computed tomography (CT) and magnetic resonance imaging (MRI) based radiomics can be used for predicting radiation induced xerostomia. HNC patients treated between 2009 and 2015 at our institution with xerostomia graded by NCI-CTCAE at 3-month follow up were included. For each patient, contours of the parotid and submandibular glands on baseline planning CT (non-contrast enhanced) were prospectively collected. CT images were corrected for metal artifact reduction to minimize the effect of dental artifacts in radiomic feature computation. Baseline MR images were acquired post-contrast using a T1-weighted pulse sequence. CT images and corresponding contours were registered to MR images using rigid or deformable registration with commercially available software. In each patient, 5744 CT and MR features were extracted for ipsilateral(IL)/contralateral(CL) salivary glands. The dataset was randomly partitioned into training and validation sets using an 80/20 split. Features from the training set were first ranked in univariate analysis by likelihood for predicting severe xerostomia (grade ≥2). Features correlated with xerostomia (p<0.1) were further reduced using a maximum relevance minimum redundancy method. The final set of image features was selected using LASSO logistic regression. A Random Forest model with leave-one-out cross-validation was applied to predict radiation induced xerostomia. Three models were developed with image features extracted from (1) CT only, (2) MR only, and (3) CT+MR. Area under the receiver operator curve (AUC) was compared across models. The training dataset (n=112) revealed 11 CT and 6 MR image features. CT features included: CL parotid Gray Level Size Zone matrix (GLSZM) gray level non-uniformity (consistent with parotid gland heterogeneity) and Gray Level Co-occurrence Matrix (GLCM) correlation; IL submandibular GLCM entropy and total energy. MR features included: CL parotid Short Run High Gray Level Emphasis and total energy features. The cross-validated AUCs for the training data for CT only/MR only/CT+MR was 0.72/0.60/0.74 in xerostomia prediction. The AUC for the validation data (n=28) for CT only/MR only/CT+MR features was 0.82/0.68/0.73. In this cohort of HNC patients, these results suggest that CT image features may be used to predispose patients susceptible to radiation induced xerostomia. Further investigation with a larger cohort is required.