Session Item

This work aims to identify imaging biomarkers to early detect osteoradionecrosis (ORN) in head and neck cancer patients after radiation treatment (RT).

This retrospective study was approved by the institutional review board. We identified patients with confirmed ORN diagnosis at MD Anderson Cancer Center between 2008 and 2018. We retrieved all the post-RT contrast-enhanced CT scans (CECTs) of these patients and selected the first study after the ORN diagnosis. First, we manually segmented the mandibular ORN volumes in all studies using a minimalistic approach (i.e. the least volume possible was segmented without including “a safety margin” of normal bone) to ensure that the image region defined by the mask contains textural features of ORN only. Then, control normal mandibular volumes were created on the contralateral healthy mandible. The segmentations on the CECTs of both ORN and control masks were used to extract the radiomic features using the PyRadiomics® platform version 2.1.1 after the application of intrinsic filters. Redundant radiomic features that are highly correlated with other features were removed when pairwise correlation≥0.99. Subsequently, filter algorithms were used to further reduce the number of radiomic features. After that, wrapper and embedded methods were applied on the resulting radiomic features. Gini importance and Recursive Feature Elimination (RFE) were used to select the final radiomic features for the predictive model. Internal validation of the classifier was done using 5-folds cross-validation (CV-5). The performance of the model was evaluated using Area Under Curve (AUC) of the Receiver Operator Curve (ROC). The workflow is detailed in Figure 1.

A total of 150 patients with radiologically established ORN were included in our study. The mean age was 62.3 years (range 27-82). The mean duration between the end of RT and ORN diagnosis was 32.6 months. The pairwise correlation omitted 432 features with a correlation ≥ 0.99. After that, the first step of the radiomic features engineering (using the filter algorithm) resulted in the selection of 33 radiomic features with statistically significant results in all the following three statistical methods: Pearson correlation, Chi-square test, and F-score. The RFE based on the Gini index selected 5 radiomics features. The final classifier used SVM with linear Kernel. The input for this classifier was the final set of radiomic features (N=5). We validated this binary classification model using 5-fold cross-validation. During this validation, the range of AUC was (0.84–0.95) & the average AUC was 0.90. (Figure 2).

We successfully using imaging radiomic features to construct an accurate model (AUC= 0.90) to discriminate ORN and normal mandibular bone in head and neck cancer patients. Future studies are needed to validate this model in prospective studies to early detect ORN in head and neck cancer patients after radiation treatment.