基于影像组学的放射治疗计划特征预测调强放疗计划复杂度的可行性研究

Feasibility study of radiomics-based radiotherapy planning characteristics to predict the complexity of intensity-modulated radiotherapy plans

目的:探讨采用机器学习的方法提取放疗计划特征预测调强放射治疗(IMRT)计划复杂度的可行性,为IMRT计划复杂度综合评价提供新方法。方法:选取2022年12月至2023年11月在南京医科大学第一附属医院放疗科进行治疗的3203例盆腔肿瘤、腹部肿瘤及头颈部肿瘤患者病例资料,所有患者均采用Monaco系统进行计划设计,分别采用Precise和Axesse加速器进行治疗。利用Python软件计算10个放疗计划复杂度评价指标,通过格式转换借助影像组学Pyradiomics工具提取放疗计划文件中的放疗计划特征。通过机器学习的数据清洗、过滤法和嵌入法选择放疗计划特征,利用梯度提升决策树(GBDT)模型分别对10个常用的计划复杂度评价指标构建相应的预测模型,采用拟合优度(R2)值评价模型预测性能,采用五折交叉验证方法检测模型的泛化能力。结果:放疗中Precise加速器与Axesse加速器在平均叶片对面积(LA)、射束形状与标准圆的偏差(PI)、子野形状和面积的变异度(MCS)和叶片运动范围平均值(LT)比较,差异有统计学意义(t=63.894、-63.678、72.582、-48.858,P<0.01)。借助Pyradiomics工具共提取到计划组学特征107个,采用过滤法筛选后剩余38个,嵌入法筛选后复杂度指标特征数为4~11个。MU加权平均射野面积(MFA)、LA、叶片对间距平均值(LGA)等在验证集中的拟合优度较好(R2>0.970);叶片间距小于某阈值20 mm比例(SAS20)在验证集中的拟合优度较差(R2=0.917)。五折交叉验证结果显示所有复杂度评价指标的预测准确率均值>90%。结论:基于影像组学方法提取到的放疗计划特征可以准确预测调强放疗计划的复杂度,有望在提高患者放疗计划个体化质量保证效率、筛选更高质量的放疗计划方面发挥更大价值。

Objective:To explore the feasibility of predicting complexity of intensity modulated radiotherapy(IMRT)plan through adopted machine learning method to extract planomics features of radiotherapy,so as to provide a new method for comprehensive evaluation of the complexity of IMRT plan.Methods:The medical case data of 3203 patients with pelvic tumor,or abdominal tumor or head and neck tumor,who admitted to The First Affiliated Hospital with Nanjing Medical University from December 2022 to November 2023,were selected.All patients adopted Monaco system to conduct design for plan,and underwent treatment on Precise and Axesse accelerators.The evaluation indicator of complexity of 10 plans was calculated by using Python software,and the planomics features in the files of radiotherapy plans were extracted through format conversion and pyradiomics tool of imaging omics.The planomics features of radiotherapy were selected through data cleaning,filtering method and embedding method of machine learning.The corresponding predictive model of the evaluation indicator of complexity of 10 common plans was respectively constructed through used Gradient Boosting Decision Tree algorithm.The goodness of fit(R2)was adopted to evaluate the prediction performance of the model,and the 5-fold cross-validation method was adopted to detect the generalization ability of the model.Results:There were statistically significant differences between Precise accelerator and Axesse accelerator in average leaf to area(LA),plan irregularity(PI)of beam shape and standard circle,modulation complexity score(MCS)of the variability between shape and area of subfield,and the advantage value of leaf travel(LT)(t=63.894,-63.678,72.582,-48.858,P<0.01),respectively.A total of 107 planomics features were extracted through pyradiomics tool,and 38 features were remained after filtering method conducted screening,and 4 to 11 features were remained after embedding method conducted screening.The goodness of fits of mean field area(MFA),LA and leaf gap average(LGA)value were better in the validation set,with R2>0.970,however the goodness of fits of the proportion of small aperture score 20 mm(SAS20)was poor in validation set,with R2=0.917.The 5-fold cross-validation results showed that the average value of prediction accuracy of all indicators of complexity was>90%.Conclusions:The extracted planomics features of radiotherapy based on radiomics method can accurately predict the complexity of IMRT plan,which are expected to play a greater role in improving the ensure efficiency of individual quality of patient,and screening radiotherapy plan with higher-quality.