Early prediction of skin toxicities during radiotherapy using optical and infrared imaging

Abstract

Approximately 90% of breast cancer radiation therapy patients experience skin toxicities that are difficult to predict. The objectives of this study are to evaluate correlation between the skin toxicity occurrences and features extracted from optical and infrared (thermal) images of skin, and develop a model for predicting skin response to radiation. Optical and infrared images of breast cancer patients' chest wall and skin toxicity assessments were acquired up until the third week following the end of treatment. The trends of colour and temperature features of the skin on the treated area were reduced and used in a machine learning model to predict skin toxicity grade. The cross-validation accuracy remained above 80% and AUC above 0.7 when reducing the input data to include a biologically effective dose of 30 Gy (equivalently first-third to first-half of the treatment). The quantitative analysis was shown to be promising for predicting skin toxicities.