鼻咽癌IMRT剂量验证中BP神经网络的应用  

Application of BP neural network in intensity modulated conformal radiotherapy dose verification of nasopharyngeal cancer

作　　者：陈龙[1,2] 高韩[2,3] 倪婕 涂彧[3] 孙亮[2,3] 秦颂兵[1] CHEN Long;GAO Han;NI Jie;TU Yu;SUN Liang;QIN Song-bing(Department of Radiation Oncology,First Affiliated Hospital of Soochow University,Suzhou 215123,China;Soochow University,Suzhou 215006,China)

机构地区：[1]苏州大学附属第一医院放疗科,江苏苏州215123 [2]苏州大学放射医学与辐射防护国家重点实验室辐射防护与核安全研究中心,江苏苏州215006 [3]苏州大学江苏省高校放射医学协同创新中心,江苏苏州215006

出　　处：《中华肿瘤防治杂志》2022年第10期736-741,共6页Chinese Journal of Cancer Prevention and Treatment

基　　金：国家自然科学基金(11575124);国家核能开发项目(2016-1295)。

摘　　要：目的使用基于反向传播(BP)神经网络的机器学习方法对鼻咽癌患者的Gamma通过率建立预测模型,并评估其实用性。方法收集2018-01-01-2020-10-31苏州大学附属第一医院放疗科收治的106例Ⅱ~Ⅲ期鼻咽癌患者临床资料,均采用同步加量9野均分动态调强治疗方案。提取患者放疗计划中与射野复杂性相关的31个指标,同时用mapcheck2测量954个射野的实际剂量分布,得到2%/2 mm、2%/3 mm、3%/2 mm和3%/3 mm(10%阈值和全局归一)4组参数下的Gamma通过率,将射野复杂度指标与4组Gamma通过率分别进行Spearman相关性分析,选择其中有统计学意义的指标作为神经网络学习的输入特征,90%数据作为训练集,10%数据作为测试集,对4组参数分别建立神经网络预测模型。结果在与射野复杂性相关的31个指标中,调制复杂度、射野调制性等多个指标与测量Gamma通过率的Spearman相关系数差异有统计学意义,均P<0.05。2%/2 mm、2%/3 mm、3%/2 mm和3%/3 mm 4组神经网络模型预测的Gamma通过率最大偏差分别为4.10%、2.80%、3.20%和2.77%,其中99%的射野预测值与实际测量值的绝对偏差≤3%。结论基于计划射野复杂性特征建立的BP神经网络模型可以对不同标准Gamma通过率进行准确预测,有助于提前发现Gamma通过率低的计划,对临床计划设计和放射治疗质量保证(QA)工作都具有一定的指导意义。Objective To establish a predictive model for the Gamma passing rates of nasopharyngeal cancer and assess its usefulness by using machine learning method based on back propogation(BP) neural networks.Methods The clinical data of 106 patients with stage Ⅱ-Ⅲ nasopharyngeal carcinoma treated in the Radiotherapy Department of First Affiliated Hospital of Soochow University from January 1,2018 to October 31,2020 were collected.All patients were treated with dynamic intensity-modulated therapy with 9 fields.31 features related to the complexity of the fields in the radiotherapy plans were extracted, and the actual dose distribution of 954 fields were measured with mapcheck2.The Gamma passing rate with 4 gamma criteria of 2%/2,2%/3,3%/2 and 3%/3 mm(10% threshold and global normalization) were obtained.The beam field complexity features and the Gamma passing rates of the four groups were analyzed separately by Spearman correlation analysis.The statistically significant features were selected as the input characteristics of the neural network learning, 90% of the data as the training set, and 10% of the data as the testing set, and the neural network prediction model was established for 4 sets of parameters.Results Among the 31 features related to the complexity of the fields, the modulation complexity score(MCS),beam modulation(BM) and other features were statistically significant with the Spearman correlation coefficient for the Gamma passing rates of Gamma(P<0.05),and the maximum deviation of the Gamma passing rates predicted by the four sets of neural network models was 4.10%,2.80%,3.20%and 2.77%,respectively,of which 99% of the absolute deviation between the predicted value of the field and the actual measurement was≤3%.Conclusion The BP neural network model based on the complexity characteristics of the planned field can accurately predict the passing rates of different standard Gamma,which helps detect the plan with low Gamma passing rates in advance and has certain guiding significance for clinical plan desi

关 键 词：剂量验证 神经网络 Gamma通过率 鼻咽癌IMRT 

分 类 号：R739.63[医药卫生—肿瘤]