综合复杂性特征和剂量学评估指标提高剂量验证结果预测模型的性能

Comprehensive index combining plan complexity characteristics and dosimetric evaluation indicators to improve model performance for predicting the results of dose verification

目的:构建随机森林模型预测调强计划剂量验证结果,研究综合复杂性特征和剂量学评估指标提高模型性能的可行性。方法:选取269例IMRT计划,共2 558个射野,采用电子射野影像系统进行剂量验证,γ通过率(2%/2 mm标准)阈值为95%,将剂量验证结果分为“通过”和“不通过”。提取计划的剂量学评估指标和射野的复杂性特征,分别构建剂量模型(基于剂量学评估指标)、计划模型(基于计划复杂性特征)和混合模型(综合剂量学评估指标和计划复杂性特征)。计算AUC值、特异性和敏感性评估模型性能。结果:剂量模型、计划模型和混合模型的AUC值分别为0.68、0.80和0.82,混合模型优于其他两个模型。混合模型的特异性和敏感性为0.70和0.79,均高于其他两个模型。剂量模型、计划模型和混合模型达到最佳性能所需的样本量分别为1200、900和700。结论:剂量学评估指标与计划复杂性特征综合,可以提高模型的预测性能,同时在一定程度上弥补样本数量的不足,为预测剂量验证结果的机器学习模型性能的改善提供参考。

Objective To develop a random forest model for predicting the results of intensity-modulated radiotherapy(IMRT)plan dose verification, and to study the feasibility of improving model performance by integrating plan complexity characteristics and dosimetric evaluation indicators. Methods Electronic portal imaging device was used for the dose verification of 269 IMRT plans with a total of 2 558 fields. The threshold of gamma passing rate(2%/2 mm criterion) was95%, and there were only two possible outcomes in dose verification, namely pass and fail. The dosimetric evaluation indicators of plans and the complexity characteristics of the radiation fields were extracted for constructing the dose model(based on dosimetric evaluation indicators), planning model(based on plan complexity characteristics) and the hybrid model(comprehensively considering dosimetric evaluation indicators and plan complexity characteristics). The performances of the prediction models were evaluated by AUC, specificity and sensitivity. Results The AUC of the dose model, the planning model and the hybrid model were 0.68, 0.80 and 0.82, respectively, and the hybrid model had the highest AUC. The specificity and sensitivity of the hybrid model were 0.70 and 0.79, both higher than those of the other two models. The number of samples required for the optimal performance of the dose model the planning model and hybrid model were 1 200,900 and 700, respectively. Conclusion Comprehensively considering dosimetric evaluation indicators and plan complexity characteristics can improve the prediction performance of the model, and at the same time make up for the lack of sample size to a certain extent, providing a reference for improvement of the performance of machine learning model for predicting the results of dose verification.