Predicting Radiation-Adverse Effects Using Three-Dimensional Dose and Fractionation Data: Radiation Dermatitis

Abstract

Predicting the radiation dose-toxicity relationship is important to patients, as iatrogenic toxicities affect patients’ quality of life. Patients need accurate estimates of both treatment efficacy and toxicity risk to make informed decisions about a proposed treatment. QUANTEC uses point dose-volume, and mean doses to predict radiation induced toxicities. These models do not account for major factors known to affect organ toxicity, such as dose per fraction, radiation field volume, and anatomic location of the dose within the organ of interest. This study evaluates utilizing 3 dimensional (3D) dose images, and fractionation data to predict radiation dermatitis in patients undergoing whole breast radiotherapy. One hundred and sixty patients underwent whole breast radiotherapy for ductal carcinoma in-situ or early stage breast cancer following breast conserving surgery. Patients were randomly separated into training or validation datasets. Three-dimensional radiation dose images and fractionation data were used to train a modified three-dimensional squeeze excitation residual neural network to predict radiation dermatitis grade. Receiver operating characteristic analysis was used to assess the model’s performance. Sixty seven patients had grade ≥ 2 radiation dermatitis, and 93 patients had grade ≤ 1 radiation dermatitis. The neural network model discriminated between radiation dermatitis grade ≤ 1 from grade ≥2 with an area under the curve (AUC) of 0.80 on the training dataset, and an AUC of 0.81 on the validation dataset. We have developed the first normal tissue complication probability model that utilizes 3D dose images and radiotherapy fractionation data to reasonably predict radiation dermatitis in patients undergoing whole breast radiotherapy. With advancements in 3D computer modelling, radiation oncologists will be able to better predict the probability and severity radiation toxicities compared to historical point dose-volume and mean dose toxicity models currently used in practice. With improved toxicity predictions, patients and physicians will be able to make more informed treatment decisions.