Comparisons of Outcome Prediction Performance between Radiomics Features and Clinical Features Based on NRG Oncology/ RTOG-0522

Abstract

This project investigated radiomics features, clinical features, and their combination for predicting clinical outcomes of head and neck carcinoma. We hypothesized that the combination of radiomics and clinical features in prediction models would lead to better prediction performance. Data were obtained from NRG RTOG 0522, i a randomized phase III trial of concurrent accelerated radiation plus cisplatin with or without cetuximab for stage III to IV head and neck carcinoma. 940 patients enrolled in the trial. A total of 773 cases with complete clinical and image data were included in this study. All the cases underwent planning CT scans before chemo-radiotherapy. A total of 107 radiomics features were extracted from the filtered images, including 14 shape features, 18 first-order features, 24 GLCM texture features, 14 GLCM texture features, 16 GLRLM texture features, 16 GLSZM texture features, and 5 NGTDM texture features. Clinical features include: age, gender, clinical TNM staging, tumor size, smoking status, Zubrod status, hemoglobin level, radiotherapy dose and fractions. Overall survival and local recurrence were observed as the outcome in this study. Univariate analysis was applied to all features for feature selection. The top 4 selected features were used to build Cox Regression models for outcome prediction. The model was validated by ten-fold cross-validation. The prediction performance of models built on radiomics features was compared with models built on clinical features and a combination of both clinical and radiomics features. The Cox regression models built on radiomics features, clinical features, and their combination achieved testing C-indices of 0.59, 0.67, and 0.68 respectively for predicting overall survival, and achieved testing C-indices of 0.58, 0.55, 0.59 respectively for predicting recurrence-free survival. For predicting overall survival, models built with clinical features had better performance than those built on radiomics features only, and the prediction model built on their combination had the overall best performance. For predicting recurrence-free survival, models built on radiomics features had better performance than those based on clinical features, and the model built on their combination improved the prediction performance.