基于计划复杂度参数利用机器学习预测容积弧形调强验证通过率研究

Predicting passing rate for VMAT validation using machine learning based on plan complexity parameters

目的采用随机森林和支持向量机两种机器学习算法建立预测模型, 实现容积弧形调强(VMAT)验证γ通过率(GPR)的数值预测和分类预测。方法回顾性选取2019年4月至2020年8月于温州医科大学附属第一医院放疗中心接受VMAT放疗的258例患者资料, 其中头颈部38例, 胸腹部220例。对所有患者VMAT计划提取13个复杂度参数, 收集计划验证在3%/3 mm和2%/2 mm评估标准下的GPR。用随机种子数方法将患者队列分为70%训练队列和30%验证队列, 采用随机森林和最小冗余最大相关性(mRMR)的方法分别筛选数值预测和分类预测的相关性参数, 利用随机森林和支持向量机两种算法结合计划靶区(PTV)体积, 子野宽度, 平滑度因素分别建立复杂度模型和混合模型。分析比较两种模型的预测性能。结果对GPR数值预测, 在3%/3 mm、2%/2 mm评估标准下, 验证队列中基于随机森林和支持向量机复杂度模型预测误差的均方根误差(RMSE)和平均绝对误差(MAE)分别为1.788%和1.753%、5.895%和5.444%、1.415%和1.334%、4.644%和4.255%;基于随机森林和支持向量机的混合模型预测误差的RMSE和MAE分别为1.760%和1.815%、5.693%和5.590%、1.386%和1.319%、4.523%和4.310%。GPR分类预测中, 在3%/3 mm、2%/2 mm评估标准下, 验证队列中基于随机森林和支持向量机的复杂度模型的受试者工作特征(ROC)曲线下面积(AUC)分别为0.790和0.793、0.763和0.754;在验证队列中基于随机森林和支持向量机的混合模型的AUC结果分别为0.806和0.859、0.796和0.796。结论开发了基于随机森林方法和支持向量机方法的复杂度模型和混合预测模型, 能够实现在3%/3 mm和2%/2 mm两种评估标准下对VMAT放疗计划GPR的数值预测和分类预测, 混合模型较经典复杂度模型提高了预测精度。

Objective To establishaprediction model using the random forest(RF)and support vector machine(SVM)algorithms to achieve the numerical and classification predictions of the gamma passing rate(GPR)for volumetric arc intensity modulation(VMAT)validation.Methods Atotal of 258 patients who received VMAT radiotherapy in the 1st Affiliated Hospital of Wenzhou Medical University from April 2019 to August 2020 were retrospectively selected for patient-specific QA measurements,including 38 patients who received VMAT radiotherapy for head and neck,and 220 patients who received VMAT radiotherapy for chest and abdomen.Thirteen complexity parameters were extracted from the patient′s VMAT plans and the GPRs for VMAT validation under the analysis criteria of 3%/3 mm and 2%/2 mm were collected.The patients were randomly divided into a training cohort(70%)and a validation cohort(30%),and the complexity parameters for the numerical and classification predictions were screened using the RF and minimum redundancy maximum correlation(mRMR)method,respectively.Complexity models and mixed models were established using PTV volume,subfield width,and smoothness factors based on the RF and SVM algorithms individually.The prediction performance of the established models was analyzed and compared.Results For the validation cohort,the GPR numerical prediction errors of the complexity models based on RF and SVM under the two analysis criteria are as follows.The root-mean-square errors(RMSEs)under the analysis criterion of 3%/3 mm were 1.788%and 1.753%,respectively;the RMSEs under the analysis criterion of 2%/2 mm were 5.895%and 5.444%,respectively;the mean absolute errors(MAEs)under the analysis criterion of 3%/3 mm were 1.415%and 1.334%,respectively,and the MAEs under the analysis criteria of 2%/2 mm were 4.644%and 4.255%,respectively.For the validation cohort,the GPR numerical prediction errors of the mixed models based on RF and SVM under the two analysis criteria were as follows.The RMSEs under the analysis criterion of 3%/3 mm were 1.760%and 1.815%,respectively;the RMSEs under the analysis criterion of 2%/2 mm were 5.693%and 5.590%,respectively;the MAEs under the analysis criterion of 3%/3 mm were 1.386%and 1.319%,respectively,and the MAEs under the analysis criteria of 2%/2 mm were 4.523%and 4.310,respectively.For the validation cohort,the AUC result of the GPR classification prediction of the complexity models based on RF and SVM were 0.790 and 0.793,respectively under the analysis criterion of 3%/3 mm and were 0.763 and 0.754,respectively under the analysis criterion of 2%/2 mm.For the validation cohort,the AUC result of the GPR classification prediction of the mixed models based on RF and SVM were 0.806 and 0.859,respectively under the analysis criterion of 3%/3 mm and were 0.796 and 0.796,respectively under the analysis criterion of 2%/2 mm cohort.Conclusions Complexity models and mixed models were developed based on the RF and SVM method.Both types of models allow for the numerical and classification predictions of the GPRs of VMAT radiotherapy plans under analysis criteria of 3%/3 mm and 2%/2 mm.The mixed models have higher prediction accuracy than the complexity models.