Prediction of acute oral mucositis and overall survival for nasopharyngeal carcinoma patients with radiation therapy using radiomics

Abstract

Nasopharyngeal carcinoma (NPC) is a malignant neoplasm that arises from the mucosal epithelium of the nasopharynx. In the field of clinical practice, radiation therapy (RT) remains the primary treatment modality for NPC patients. However, with the increasing emphasis on personalized treatment approaches and advancements in artificial intelligence (AI), researchers have started utilizing AI-based cancer imaging analysis, which combines clinical images with AI techniques, to enhance various clinical tasks related to NPC patients. Predicting outcomes, such as overall survival (OS) and incidence of toxicity after RT, for NPC patients can assist clinicians in assessing the risk profile based on tumor characteristics. This information permits identification of patients with poor prognosis, who may benefit from escalated therapy or inclusion in clinical trials. Conversely, patients with a favorable prognosis, if identified in advance, could receive de-escalated therapy, minimizing the physiological and financial burdens associated with cancer treatment. However, toxicity predictions, specifically for the incidence of acute oral mucositis (AOM), are mainly using single sources of data and has limited predictive capability of models. Similarly, studies related to survival prediction models often lack external validation, especially international validation, calling for the development of more generalizable models. The aims in our studies are focused mainly on two parts, to investigate the impact of radiomics, dosiomics, extracted from multi-regions and multi-sources, and clinical data on the prediction of severe acute oral mucositis in patients undergoing radiotherapy for NPC, and to develop CT-based generalizable prognostic model with perturbation in an international dataset for the prediction of five-year OS of NPC patients following intensity-modulated radiation therapy (IMRT). For the prediction of AOM, pathological validated NPC patients were retrospectively included from Queen Elizabeth Hospital (QEH) in Hong Kong. Radiomics features (RFs) and dosiomics features were extracted from various volume of interests (VOIs): Gross tumor volume of primary NPC tumor (GTVp), metastatic lymph nodes area (GTVn), regions of nodal planning target volume with the prescribed dose level of 70Gy (PTVn_70Gy), PTVn with and the prescribed dose level of 60Gy (PTVn_60Gy), using contrast enhanced computed tomography (CECT), cT1-weighted imaging (cT1WI), T2-weighted imaging (T2WI), and dose-volume histogram (DVH) data. Additionally, relevant clinical variables were incorporated into the analysis. Logistic Regression (LR), Gaussian Naive Bayes (GNB) and eXtreme Gradient Boosting (XGBoost) models were developed, considering different combinations of data extracted from distinct VOIs and image modalities. The area under the curve (AUC) of the receiver operating characteristic (ROC) was used to assess the performance of the models. For the prediction of OS, patients were sourced from both a private database in Hong Kong and the publicly available RADCURE database in Canada. RFs were extracted from GTVp in computed tomography (CT) images. Perturbations of the images were employed to select robust RFs. Conventional machine learning models and a multilayer perceptron (MLP) model were developed, integrating the RFs with clinical variables to predict the 5-year OS of NPC patients. The AUC of the ROC was used as a metric to evaluate the performance of the models. For the first study, the best performing GNB model with 10-fold cross validation AUC of 0.81±0.1 was developed for the prediction of AOM with radiomics and dosiomics features extracted from primary tumor area. For the second study, the best performing MLP score/LR model achieved an internal validation AUC of 0.734 [95% confidence interval (95% CI): 0.765-0.865] and an external validation AUC of 0.735 (95% CI: 0.681-0.783). Our studies have shown that clinical images hold promise for predicting the occurrence of AOM and OS in NPC patients. In particular, utilizing multimodal data sources has the potential to enhance the performance of the prediction model for AOM. Additionally, deep learning model have demonstrated its effectiveness in handling data from various institutions. These findings provide valuable insights and lay the groundwork for further research in predicting outcomes for NPC patients, particularly in the context of patient screening.