The Role of Radiomics Model to Predict Early Xerostomia in Head and Neck Cancer after Radiotherapy

Abstract

Based on the patients' characteristics, the most treated patients suffering from early xerostomia during radiation therapy treatment and in the early months after treatment in head and neck cancer. The aim of this study is to provide an image-based model to predict early xerostomia in these patients before radiotherapy. The dosimetric and clinical factors (age, sex, history of disease) and computed tomography (CT) features are also combined with the image data in resolving xerostomia detection. Clinical factors and CT features were considered for 62 patients with head and neck cancer, and the complication of parotid and salivary glands was evaluated. For feature extraction from CT images, 3D Slicer software was used to extract 107 features. The features type included shape, gray level dependence matrix, gray level co-occurrence matrix, first order gray level run length matrix, gray level size zone matrix, and neighborhood gray-tone difference matrix. Four different datasets; demographic + sum dose volume histogram (DVH), demographic + ipsilateral and contralateral DVH, demographic + CT features + sum DVH, and demographic + CT features + ipsilateral and contralateral DVH; were used as model inputs. Four algorithms: Random Tree (RT), Linear Support Vector Machine (LSVM), Neural Network (NN), and Bayesian Network (BN), were applied to analyze the datasets to predict xerostomia. The metrics of the models used sensitivity, specificity, accuracy, and area-under-the-curve (AUC) to assess the performance. The AUC has its highest value for the RT model among the aforementioned models and its value for the first and second data sets were 0.84 and 0.64, respectively. In general, model performance was improved when the sum of the parotids DVH data was used. Furthermore, the RT model showed the best performance when using radiomics features in the third and fourth data sets with the AUC of 0.86 and 0.89, respectively. Among the RT, LSVM, NN, and BN models, RT has the highest performance in predicting radiation-induced early xerostomia in patients who underwent head and neck radiotherapy. Model performance was improved when the sum of the parotids DVH data was used in addition to the dose data and CT features. This study showed that combining the dosimetric data with radiomics features may improve the model performance for predicting early xerostomia.