Predicting late radiation-induced xerostomia in nasopharyngeal carcinoma based on radiomics features extracted from T2WI images of parotids

Abstract

ObjectiveTo explore the value of radiomics features extracted from the T2-weighted imaging (T2WI) images of parotids in predicting late radiation-induced xerostomia in nasopharyngeal carcinoma (NPC) patients after radiotherapy (RT). MethodsA retrospective analysis was conducted for 123 NPC patients who received RT at the Affiliated Tumour Hospital of Xinjiang Medical University from January 2019 to March 2021. All the patients underwent MRI pre-RT and post-RT. They were randomly divided into a training set and a testing set at a ratio of 4:1 using a random number table, with the former and the latter comprising 98 and 25 cases, respectively. The ipsilateral parotid gland (iPG) and the contralateral parotid gland (cPG) were delineated on T2WI images pre-RT and post-RT as regions of interest (ROIs). A total of 851 radiomics features were extracted from each ROI. Spearman analysis was used to remove redundant features, and the recursive feature elimination (RFE) method was then used to determine useful features. Using radiomics features extracted from images pre-treatment, images post-treatment, and differences between images pre- and post-treatment, this study constructed three radiomic models, namely the pre-treatment radiomics model (preRT), the post-treatment radiomics model (postRT), and the delta-radiomics model (DeltaRT). Then, this study plotted the receiver operating characteristic (ROC) curves based on the late radiation-induced xerostomia grades of patients post-RT. Furthermore, the models’ effectiveness and performance in predicting late radiation-induced xerostomia and advanced radioactive xerostomia was evaluated. In addition, the area under the curve (AUC), sensitivity, specificity, accuracy, precision, and negative predictive value (NPV) were calculated. ResultsAmong the features extracted from bilateral parotid glands (PGs), 20 were determined pre-RT (six from iPG and 14 from cPG), 19 were determined post-RT (six from iPG and 13 from cPG), and 20 were derived from the DeltaRT (20 from cPG). The PGs pre-RT and post-RT in the training set had AUCs of 0.902 (95% CI: 0.895–0.909) and 0.761 (95% CI: 0.744–0.778), respectively, while those in the testing set had AUCs of 0.740 (95% CI: 0.504–0.983) and 0.701 (95% CI: 0.478–0.924), respectively. In contrast, the AUC of the cPG derived from the DeltaRT was 0.867 (95% CI: 0.856–0.878) in the training set and 0.851 (95% CI: 0.697–0.999) in the testing set. ConclusionsThere are significant correlations between radiomics features extracted from MRI T2WI images of parotids and late radiation-induced xerostomia in NPC patients. Among the radiomics features, the changes in cPG features pre-RT and post-RT have higher accuracy in predicting late radiation-induced xerostomia.