Radiation therapy is an essential part of the comprehensive breast cancer treatment strategy, and radiation dermatitis is the inevitable side-effect. According to either patient-related or treatment-related factors, patients will experience different degrees of acute radiation dermatitis. This study proposes a machine learning architecture based on image and time series features. Using the skin image of the irradiated part during radiotherapy, the image feature is extracted with a gray-level co-occurrence matrix (GLCM) and color space, combined with the time series feature with gradient boosting decision trees (GBDT) to predict the severity of dermatitis after seven days of treatment. The results show that, through the combination of image and time series features, the predicted accuracy (ACC) and area under the curve (AUC) can be effectively improved to 0.8 and 0.85 respectively. The results of GBDT show higher prediction accuracy and robustness than AdaBoost algorithm. This framework can be used as an auxiliary diagnostic tool to assist doctors in making appropriate treatments before severe dermatitis occurs, in order to reduce the radiotoxicity caused by radiotherapy of patients.

中文翻译：

---

使用机器学习算法预测乳腺癌患者放射性皮炎的照片图像处理

放射治疗是乳腺癌综合治疗策略的重要组成部分，放射性皮炎是不可避免的副作用。根据患者相关因素或治疗相关因素，患者会出现不同程度的急性放射性皮炎。本研究提出了一种基于图像和时间序列特征的机器学习架构。利用放疗过程中受照射部位的皮肤图像，利用灰度共生矩阵（GLCM）和色彩空间提取图像特征，结合梯度提升决策树（GBDT）的时间序列特征预测严重程度治疗 7 天后出现皮炎。结果表明，通过图像和时间序列特征的结合，预测准确率（ACC）和曲线下面积（AUC）可以分别有效提高到0.8和0.85。GBDT的结果显示出比AdaBoost算法更高的预测精度和鲁棒性。该框架可作为辅助诊断工具，协助医生在严重皮炎发生前进行适当的治疗，以减少患者放射治疗引起的放射毒性。