Radiomic Analysis of Primary GTV and CTV for Prediction of Extranodal Extension Using Diagnostic CT Images in Patients With Oropharyngeal Squamous Cell Carcinoma

Document Type

Conference Proceeding

Publication Date


Publication Title

Int J Radiat Oncol Biol Phys


Purpose/Objective(s): Prediction of extranodal extension (ENE) status of the involved lymph nodes has been shown to be an important factor in the prognosis and treatment of patients with Oropharyngeal Squamous Cell Carcinomas (OPSCC). Official diagnosis of ENE occurs pathologically at the time of surgery, and typically necessitates aggressive adjuvant treatment with radiation and concurrent chemotherapy. Due to higher toxicity with trimodality therapy, definitive chemoradiation can be a preferred treatment option up front instead of surgery if there is high clinical suspicion of ENE. This study investigates the potential value of radiomic features extracted from pre-treatment primary gross tumor volume (GTV) and clinical target volume (CTV) on diagnostic contrast enhanced Computed Tomography (CE-CT) images for prediction of ENE in patients with OPSCC.

Materials/Methods: For this study we identified 26 patients with pathologically confirmed OPSCC who went on to receive surgery and neck nodal dissection. Out of this group 13 were p16 positive, 13 were p16 negative, 7 were pathologic ENE positive, and 19 pathologic ENE negative. We contoured the primary tumor GTV on their pre-surgery CE-CT and added a 1 cm to create the CTV for radiomic analysis. For each patient, two sets of IBSI (Image Biomarker Standardization Initiative) validated radiomic features (N = 192 features from 11 different feature categories) were extracted from primary GTV and CTV on the diagnostic CE-CT images. Levene and Kolmogorov-Smirnov's (KS, P < 0.05) tests with confidence level of 95% along with absolute biserial correlation (|BSC|) with a threshold of > 0.2 were used to statistically reveal significant associations between ENE status and radiomics features. The Belsley collinearity diagnostics test was applied on the discriminant radiomic features for removing highly correlated features. Eight different Generalized-Linear Models (GLMs) were trained and tested using each individual feature along with two combined feature sets (combined discriminant features extracted from GTV and CTV). Balanced Accuracy Score (BAS), Positive predictive and negative predictive values (PPV and NPV, respectively) were used to evaluate the performance of the classifiers.

Results: Four GTV-based radiomic features and two CTV-based radiomic features (both from textural feature categories) were found to be statistically significant discriminators between the two ENE cohorts. Performances for prediction of the ENE for the two classifiers trained with the combined feature sets were: GTV-based GLM: BAS/PPV/NPV = 0.765/0.724/0.805 and CTV-based GLM: 0.875/0.928/0.822. Results imply that the CTV-based features have greater predictive value compared to GTV-based features for characterization of ENE.

Conclusion: This pilot study, albeit subject to confirmation in a larger patient population, suggests potential for the use of radiomics-based signatures extracted from the primary tumor for prediction of ENE in patients with OPSCC.





First Page