基于国产加速器URT-Linac 506c的计划复杂度分析和患者计划质量保证预测

Analysis of Complexity Metrics and Patient-Specific Quality Assurance Prediction Based on Domestically Made Accelerator URT-Linac 506c

目的 探讨基于国产联影URT-Linac 506c加速器的容积弧形调强放射治疗(Volumetric Modulated Arc Therapy,VMAT)计划复杂度参数与患者计划质量保证(Patient Specific Quality Assurance,PSQA)间的关系,并建立机器学习模型对PSQA结果进行评估和预测。方法 随机选取150例在URT-Linac 506c加速器上行VMAT治疗计划的患者为研究对象,在该加速器上对所有计划进行基于机载电子射野影像系统探测器的PSQA剂量验证。对剂量验证结果进行阈值为10%、标准为2%/2 mm的伽马通过率分析。同时对每个计划基于多叶准直器位置和跳数,提取其中的11个复杂度参数。分析复杂度参数和PSQA结果间的关系,并建立机器学习模型对PSQA结果进行预测。结果 计划复杂度参数与PSQA结果的相关性分析表明二者并不严格呈线性相关,但计划的复杂程度越高,PSQA通过率相对越低。梯度提升决策树(Gradient Boosting Decision Tree,GBDT)模型和随机森林(Random Forest,RF)模型对基于复杂度参数的PSQA结果预测水平相当,GBDT模型和RF模型的平均预测误差分别为0.55%和0.54%。由于PSQA结果分布严重不平衡,更改过权重后的模型对低通过率部分的预测能力有所提升。结论 对于国产加速器URT-Linac 506c,这2种基于决策树结构的机器学习模型对PSQA结果的预测可提供一定的帮助,建立更精准的模型需要进一步完善采集患者的样本量。

Objective To study the relationship between the complexity metrics of volumetric modulated arc therapy(VMAT)plan and patient specific quality assurance(PSQA)based on the URT-Linac 506c accelerator.And to establish the machine learning models to predict and evaluate the PSQA results.Methods A total of 150 patients treated on URT-Linac 506c accelerator in the VMAT program were randomly selected as the research object,and all plans were verified by PSQA dose based on electronic portal imaging device detector on the accelerator.The gamma pass rate of dose verification results was analyzed with the threshold of 10%and the standard of 2%/2 mm.At the same time,based on multi-leaf collimator location and monitor unit,11 complexity parameters were extracted for each plan.The relationship between complexity metrics and the results from PSQA were studied,in the meantime,two tree-based machine learning models were established to predict PSQA results.Results The correlation analysis between plan complexity metrics and the PSQA results showed that they were not strictly linearly correlated,but the higher the complexity of the plan,the lower the PSQA pass rate.The gradient boosting decision tree(GBDT)model and random forest(RF)model had similar prediction level for PSQA with average prediction errors of were 0.55%of GBDT and 0.54%of RF respectively.Due to the imbalance in the distribution of PSQA results,the model with changed weights indeed could improve the prediction ability for plans with low pass rates.Conclusion For the domestically-made accelerator URT-Linac 506c,the two tree-based machine learning models showed in this study can provide certain assistance in predicting PSQA results.The establishment of a more accurate model needs to further improve the sample size of the collected patients.