机器学习在肺癌VMAT计划中对危及器官剂量预测的可行性

目的:探讨机器学习在肺癌容积旋转调强（VMAT）治疗计划对心脏和肺的剂量体积直方图（DVH）预测的可行性。方法:选取51例肺癌VMAT计划，随机选取其中43例为训练组，剩余8例为验证组。分析训练组中患者的解剖信息与两侧肺V5、V20和心脏V30、V40的相关性。采用机器学习方法，以解剖信息为输入、危及器官（OAR）的DVH为输出，分别构建并训练关于两侧肺以及心脏的人工神经网络模型。将验证组中8例VMAT计划中的解剖信息分别输入到已经构建好的人工神经网络模型，分别预测OAR的DVH。结果:两侧肺V5、V20和心脏V30、V40受自身体积大小影响可忽略，受OAR与靶区的空间相对位置关系影响较大。患侧肺、对侧肺、心脏的人工神经网络结构模型中隐藏层分别含有41、38、34个神经结点，线性回归系数分别为0.994、0.975、0.986。对验证组中患侧肺和对侧肺的V5、V20的预测误差分别为2.70%±]1.83%、2.84%±]1.97%和13.7%±]7.8%、0.72%±]0.75%，对心脏V30、V40的预测误差分别为3.20%±]0.63%、2.1%±]1.5%，仅对侧肺V5的预测值和实际值差异有统计学意义（P<0.05）。结论:采用人工神经网络方法可以对肺癌VMAT计划中解剖信息与OAR的DVH数据进行学习，构建的人工神经网络模型可预测出患侧肺、心脏V25~]V60和对侧肺V20的DVH数据，可为临床计划设计提供参考。

Feasibility of machine learning in OAR dosimetric prediction in VMAT plan for lung cancer

Abstract: Objective To investigate the feasibility of machine learning for dose-volume histogram (DVH) predictions of the heart and the lungs in volumetric modulated arc therapy (VMAT) plan for lung cancer. Methods Among the VMAT plans of 51 cases of lung cancer, 43 VMAT plans were randomly selected as training group, and the other 8 plans were taken as validation group. The anatomical information of patients in training group was analyzed, and the relationships between the V5, V20 of bilateral lungs and the V30, V40 of the heart were investigated. With the anatomical information as the input and the DVH of organs-at-risk (OAR) as the output, machine learning method was adopted to construct and train the artificial neural network models for bilateral lungs and the heart, separately. The anatomical information of 8 VMAT plans in validation group was input into the constructed artificial neural network model for predicting the DVH of OAR. Results The V5, V20 of bilateral lungs and the V30, V40 of the heart were affected by the relative spatial relationship between OAR and target areas, but didnt affected by the volume of OAR itself. In the artificial neural network structure models of the affected lung, the contralateral lung and the heart, the hidden layers contained 41, 38 and 34 neural nodes, respectively, and the linear regression coefficients were 0.994, 0.975 and 0.986, respectively. In validation group, the prediction errors for the V5, V20 of the affected lung were 2.70%±1.83% and 2.84%±1.97%, and those for the V5, V20 of the contralateral lung were 13.7%±7.8% and 0.72%±0.75% and the prediction errors for the V30 and V40 of the heart were 3.20%±0.63% and 2.1%±1.5%, respectively. There was statistically significant difference between the predicted and actual values of the V5 of the contralateral lung. Conclusion Artificial neural network method can learn the anatomical information in the lung cancer VMAT plan and the DVH data of OAR. The constructed artificial neural network model can be used to accurately predict the DVH of the affected lung, the V25-V60 of the heart and the V20 of the contralateral lung, providing reference for clinical treatment planning.