Prediction of treatment outcome using MRI radiomics and machine learning in oropharyngeal cancer patients after surgical treatment

Abstract

Objectives: In this study, we aimed to analyze preoperative MRI images of oropharyngeal cancer patients who underwent surgical treatment, extracted radiomics features, and constructed a disease recurrence and death prediction model using radiomics features and machine-learning techniques.

Materials and methods: A total of 157 patients participated in this study, and 107 stable radiomics features were selected and used for constructing a predictive model.

Results: The performance of the combined model (clinical and radiomics) yielded the following results: AUC of 0.786, accuracy of 0.854, precision of 0.429, recall of 0.500, and f1 score of 0.462. The combined model showed better performance than either the clinical and radiomics only models for predicting disease recurrence. For predicting death, the combined model performance has an AUC of 0.841, accuracy of 0.771, precision of 0.308, recall of 0.667, and f1 score of 0.421. The combined model showed superior performance over the predictive model using only clinical variables. A Cox proportional hazard model using the combined variables for predicting patient death yielded a c-index value that was significantly better than that of the model including only clinical variables.

Conclusions: A predictive model using clinical variables and MRI radiomics features showed excellent performance in predicting disease recurrence and death in oropharyngeal cancer patients. In the future, a multicenter study is necessary to verify the model's performance and confirm its clinical usefulness.