The Prediction of Mandibular Osteoradionecrosis (ORN) in Head and Neck Radiotherapy Using CT-Derived Radiomic Features

Abstract

The overall lifelong risk of mandibular Osteoradionecrosis (MORN) in head and neck (HN) radiotherapy (RT) is 4-8% which may result in poor health-related quality of life (QOL). Most of these patients (70-94% of cases) developed within the first 3 years after radiotherapy. The most effective way to limit MORN is to reduce mandibular volumes receiving high RT doses. However, this strategy may also result in a reduced dose to the tumor due to its proximity to the mandible, and therefore must be restricted to patients deemed most vulnerable to MORN. The objective of this study was to design a machine learning model to predict the probability of MORN prior to HN RT. Our models were based on CT-derived radiomic features extracted from mandible contours, along with the patient's clinical features. We hypothesized that these features are related to MORN and incorporating them into a prediction model will help to identify patients at risk of MORN.Patient data was retrospectively collected from the Princess Margaret Cancer Centre, University Health Network and based on our inclusion criteria. Radiomic features were then extracted from the mandible for each patient. Patients' records have also been reviewed to collect previously published clinical risk factors called demographic-clinical (DMC) such as: prescribed radiation dose, ECOG status, smoking, drinking, TNM status, stage, chemotherapy administration, Human papillomavirus (HPV) infection, the occurrence of loco-regional failure and distant metastasis. Finally, a Random Forest (RF) classifier and three regressors (Deep Survival, Random Survival Forest (RSF) and Cox), were independently trained on three sets of features (radiomic, DMC and combined) to predict the patient's risk of ORN in 3 years after radiotherapy. Applying regressors, we were able to incorporate the time to ORN for the modelling and investigate the usefulness of having such information.In total, we analyzed CT images from 223 OPC patients with known MORN status (149 negative, 74 positive). We extracted a total of 1874 radiomic features from the segmented mandible for each patient. The best result obtained with Random Forest classifier (AUC = 0.86, 95% CI, P-value < 0.001) when it was trained upon combined features (Radiomics and DMC). ECOG, T status, smoking and HPV status are the top DMC features by mRMRe method. Among radiomic features, first order statistic features were mostly informative especially when images have been filtered by wavelet transform. However, our best regression model (RSF, c-index = 0.74, P-value = 0.01) were inferior to RF model that reflect non-efficiency of incorporating the time to ORN for the prediction.Our results showed the ability of multivariable Random Forest model in the prediction of MORN when radiomic features combined with DCM parameters. These results have the potential to assist us to detect patients vulnerable to developing MORN and adjust the operational strategies before treatment.