摘要
目的:探讨人工神经网络方法利用加速器日志文件预测动态调强放疗(dIMRT)计划多叶光栅位置偏差。方法:选择瓦里安Trilogy 50例dIMRT动态日志文件,TrueBeam和Edge各30例dIMRT轨迹日志文件。日志提取叶片计划位置、剂量、机架角度、铅门位置、叶片间距、叶片速度及运动参数等14个特征为输入,日志记录叶片实际位置为输出,隐藏层为22个神经元的3层神经网络模型。训练集包括70%数据,验证集和测试集包括30%数据,均方误差代价函数评估模型性能。结果:叶片速度是输入特征与位置偏差最相关特征,皮尔逊相关系数0.7以上。运动和静息叶片位置平均绝对偏差存在显著性差别。测试集Trilogy、TrueBeam和EDGE预测叶片位置最大均方误差分别小于9×10^(-5)、3×10^(-5)和3×10^(-5) mm^(2),预测位置与实际位置极为接近。模型预测当前叶片和其它序号叶片位置平均绝对偏差具有显著性差异(P<0.001)。结论:基于加速器日志文件人工神经网络模型可以对瓦里安dIMRT计划叶片位置进行预测。
Objective To predict the multi-leaf collimator(MLC)leaf positional deviations during dynamic intensitymodulated radiotherapy(dIMRT)by log file analysis and artificial neural network(ANN)method.Methods The dIMRT DynaLog files from Varian Trilogy(50 cases)and the dIMRT Trajectory log files from TrueBeam and Edge(30 cases of each)were retrieved.A 3-layer ANN model with hidden layer of 22 neurons was developed for each MLC leaf,with 14 features extracted from the log files,such as leaf planned positions,dose,gantry angle,jaw position,leaf gap,leaf velocity and leaf motion status,as input parameters,and the delivered leaf position recorded by log files as output.The proposed model was trained on 70%,validated and tested on 30%of the total data.Mean square error(MSE)was taken as the cost function to evaluate the model performance.Results The leaf velocity was the most relevant input feature to positional deviation,with a Pearson correlation coefficient greater than 0.7.Significant differences were found in the mean absolute error(MAE)between moving and resting leaf.The maximum MSE in predicting the leaf positions on test set were less than 9×10^(-5),3×10^(-5) and 3×10^(-5) mm^(2) for Trilogy,TrueBeam and EDGE,respectively;and the predicted leaf positions closely matched the actual positions during the treatment delivery.There was significant difference in MAE of the model in predicting the positions of current leaf and other leaves(P<0.001).Conclusion The proposed log file-based ANN model is capable of predicting the Varian MLC leaf position during dIMRT.
作者
张丽媛
贾楠
郑晓娜
张博
李松丽
韩全乡
ZHANG Liyuan;JIA Nan;ZHENG Xiaona;ZHANG Bo;LI Songli;HAN Quanxiang(Department of Radiation Oncology,Zhengzhou Central Hospital,Zhengzhou 450052,China)
出处
《中国医学物理学杂志》
CSCD
2021年第12期1495-1501,共7页
Chinese Journal of Medical Physics
基金
河南省高等学校重点科研项目(21A320066)。
关键词
动态调强放疗
多叶准直器
日志文件
人工神经网络
位置偏差
dynamic intensity-modulated radiotherapy
multi-leaf collimator
log file
artificial neural network
positional deviation
