Abstract
BACKGROUND: Predicting radiation dermatitis (RD), a common radiotherapy toxicity, is essential for clinical decision-making regarding toxicity management. This prospective study aimed to develop and validate a machine-learning model to predict the occurrence of grade ≥ 2 RD using thermal imaging in the early stages of radiotherapy in head and neck cancer. METHODS: Thermal images of neck skin surfaces were acquired weekly during radiotherapy. A total of 202 thermal images were used to calculate the difference map of neck skin temperature and analyze to extract thermal imaging features. Changes in imaging features during treatment were assessed in the two RD groups, grade ≥ 2 and grade ≤ 1 RD, classified according to the Common Terminology Criteria for Adverse Events (CTCAE) guidelines. Feature importance analysis was performed to select thermal imaging features correlated with grade ≥ 2 RD. A predictive model for grade ≥ 2 RD occurrence was developed using a machine learning algorithm and cross-validated. Area under the receiver-operating characteristic curve (AUC), precision, and sensitivity were used as evaluation metrics. RESULTS: Of the 202 thermal images, 54 images taken before the occurrence of grade ≥ 2 RD were used to develop the predictive model. Thermal radiomics features related to the homogeneity of image texture were selected as input features of the machine learning model. The gradient boosting decision tree showed an AUC of 0.84, precision of 0.70, and sensitivity of 0.75 in models trained using thermal features acquired before skin dose < 10 Gy. The support vector machine achieved a mean AUC of 0.71, precision of 0.68, and sensitivity of 0.70 for predicting grade ≥ 2 RD using thermal images obtained in the skin dose range of 10-20 Gy. CONCLUSIONS: Thermal images acquired from patients undergoing radiotherapy for head and neck cancer can be used as an early predictor of grade ≥ 2 RD and may aid in decision support for the management of acute skin toxicity from radiotherapy. However, our results should be interpreted with caution, given the limitations of this study.