Publikationsrepositorium - Helmholtz-Zentrum Dresden-Rossendorf

Purpose/Objective: Radiochemotherapy (RCT) for patients with head and neck squamous cell carcinoma (HNSCC) frequently causes xerostomia and dysphagia, which may be alleviated by treatment adaption, e.g., modulation of dose distribution to the salivary glands. Current clinical models, which are based on dosimetric parameters, mostly achieve moderate prediction accuracy. Therefore, we aimed to improve the prediction of xerostomia and dysphagia by using additional imaging biomarkers based on computed tomography (CT) scans.

Material/Methods: In this study 46 patients with UICC stage III/IV advanced head and neck squamous cell carcinoma (HNSCC) were considered (NCT00180180, [1]). All patients received primary RCT and underwent a pre-treatment CT scan without intravenous contrast agent. Patient-reported xerostomia and dysphagia were evaluated at baseline, every week during RCT, four weeks after treatment and three monthly thereafter. 5040 imaging features were extracted from the parotid and submandibular glands. Feature reproducibility tests based on the RIDER re-test data set [2] were performed leading to 1513 imaging features in total. The most informative features were selected by a univariate logistic regression analysis. The developed radiomic signature was used to train and validate multivariate logistic regression and random forest models using repeated 5-fold cross validation. The predication accuracy was assessed by the area under the curve (AUC).

Results: The logistic regression and the random forest model achieved similar performance in predicting xerostomia (AUC=0.71). The developed signature consisted of one dosimetric parameter and one imaging feature. For the prediction of dysphagia both models achieved only a moderate prediction accuracy (AUC=0.55).

Conclusions: For prediction of xerostomia, a signature was developed and showed a good performance. For dysphagia only moderately performing models could be obtained in this cohort. Based on our results, subgroups of patients at a high risk of xerostomia may be identified and offered treatment adaption. However, further investigations are currently ongoing, i.e., externally validating the developed signature, which is an important step in developing clinically relevant prediction models.