基于计划文件的机器学习模型预测胸部IMRT计划的γ通过率  

Prediction of gamma pass rate for thoracic intensity-modulated radiotherapy plan dose verification using a machine learning model based on planomics

作　　者：崔甜甜 刘湘月 孟楠 王永强 葛红 娄朝阳 李兵 Cui Tiantian;Liu Xiangyue;Meng Nan;Wang Yongqiang;Ge Hong;Lou Zhaoyang;Li Bing(Department of Radiation Oncology,Affiliated Cancer Hospital of Zhengzhou University&Henan Cancer Hospital,Zhengzhou450008,China)

机构地区：[1]郑州大学附属肿瘤医院,河南省肿瘤医院放疗科,郑州450008

出　　处：《中华放射肿瘤学杂志》2025年第1期81-87,共7页Chinese Journal of Radiation Oncology

基　　金：河南省留学人员科研择优资助和创业启动支持项目(37);河南省重大科技专项项目(221100310100);河南省重点研发与推广专项(科技攻关)项目(232102310125);河南省自然科学基金面上基金项目(242300421672)。

摘　　要：目的构建基于计划参数文件特征的机器学习分类预测模型,预测胸部肿瘤固定野调强放疗剂量验证的γ通过率,并评估计划参数文件在放射治疗质量保证中的应用。方法回顾性分析2022年8月至2023年3月河南省肿瘤医院放疗科收治的240例胸部肿瘤患者的固定野调强放疗计划,所有计划的验证均使用瓦里安加速器机载电子射野影像系统探测器采集剂量,并通过Eclipse放射治疗计划的Portal Dosimetry进行剂量验证结果的分析,γ通过率的标准为2%/2 mm、10%剂量阈值。从计划文件提取48个传统计划特征、2476个计划参数文件特征以及二者的组合特征,然后构建自编码器分类模型。计算受试者操作特征曲线下面积(AUC)值与准确度评估传统特征、计划参数文件特征以及二者组合特征在预测γ通过率方面的分类性能。通过20次训练-测试随机拆分系统性评估三种特征的预测效能。结果从提取的传统计划特征和计划参数文件特征中,选取了2个传统特征和16个计划参数文件特征。在测试集,模型使用组合特征、计划参数文件特征、传统计划特征的AUC分别为0.802±0.030、0.740±0.069、0.673±0.083,在训练集中分别为0.844±0.074、0.816±0.047、0.687±0.036。在测试集,模型使用组合特征、计划参数文件特征、传统计划特征的准确度分别为0.752±0.083、0.703±0.110、0.648±0.081,在训练集中分别为0.753±0.098、0.751±0.075、0.624±0.054。结论对于胸部固定野调强放疗计划,可以采用基于计划参数文件特征的机器学习方法构建预测γ通过率的分类模型。将计划参数文件特征和传统计划特征组合可以提高分类模型的预测性能。Objective To construct a machine learning classification prediction model using planning-omics(planomics)features to predict theγpass rate of intensity-modulated radiotherapy(IMRT)plan dose verification in fixed-field thoracic tumors,and evaluate the application of planomics in radiotherapy quality assurance.MethodsThe fixed-field IMRT plans of 240 patients with chest tumors admitted to Department of Radiotherapy,Henan Cancer Hospital from August 2022 to March 2023 were retrospectively analyzed.All plans underwent dose verification using the electronic portal imaging system detector on the Varian accelerator to collect field dose data.The dose verification results were analyzed through Portal Dosimetry in the treatment planning system of Eclipse.Theγpass rate standard was set at 2%/2 mm with a 10%dose threshold.From the planning documents,48 conventional planning features,2476 planomics features,and the combination of the previous two feature sets were extracted.Subsequently,an auto-encoder classification model was constructed.To evaluate the classification efficacy of various feature sets,20 random train-test divisions were conducted by calculating the area under the receiver operating characteristic curve(AUC)values along with the accuracy rates.ResultsAfter the feature selection,2 conventional features and 16 planomics features were finally selected.In the testing set,the AUC values for the model using combined features,planomics features,and conventional planned features were 0.802±0.030,0.740±0.069,and 0.673±0.083,respectively.In contrast,in the training set,these AUC values were 0.844±0.074,0.816±0.047,and 0.687±0.036,respectively.The accuracy rates were 0.752±0.083,0.703±0.110,and 0.648±0.081 in the testing set,and 0.753±0.098,0.751±0.075,and 0.624±0.054 in the training set for the combined,planomics,and conventional planning feature sets,respectively.ConclusionsFor thoracic fixed-field adjusted radiotherapy planning,the machine learning method based on planomics features can be utilized to b

关 键 词：机器学习 计划参数文件 调强放疗计划 放射疗法 

分 类 号：R730.55[医药卫生—肿瘤]