IMRT QA中基于后融合卷积神经网络的MLC误差分类预测  

MLC error detection using a late-fusion convolutional neural network in quality assurance for IMRT

作　　者：闫冰 余枭 王帅[1] 吴爱林[2] 张红雁 吴爱东 YAN Bing;YU Xiao;WANG Shuai;WU Ailin;ZHANG Hongyan;WU Aidong(Department of Radiation Oncology,the First Affiliated Hospital of University of Science and Technology of China(Anhui Provincial Hospital),Hefei 230001,China;Department of Radiation Oncology,West Branch of the First Affiliated Hospital of University of Science and Technology of China(Anhui Cancer Hospital),Hefei 230001,China)

机构地区：[1]中国科学技术大学附属第一医院(安徽省立医院)放疗科,安徽合肥230001 [2]中国科学技术大学附属第一医院西区(安徽省肿瘤医院)放疗科,安徽合肥230001

出　　处：《中国医学物理学杂志》2023年第8期925-932,共8页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(11805198)。

摘　　要：目的:评估基于剂量差图与Gamma分布图的多模态多通路卷积神经网络用于分类调强放射治疗(IMRT)质量保证(QA)中多叶准直器(MLC)误差的可行性及优势。方法:首先通过修改选取的68例IMRT放疗计划原始无误差照射野的MLC叶片位置用于模拟4种误差类型:平移误差、外扩误差、内收误差、随机误差,并将原始无误差计划及4种引入MLC误差计划重新导入TPS,计算PTW 729模体中的剂量分布;其次从测量和计算的剂量分布中创建剂量差图和两种通过率标准下的Gamma图作为数据集建立并训练多模态多通路卷积神经网络,其中330个剂量误差图和660个Gamma图用于测试集,其余数据集按照五折交叉验证划分为训练集与验证集。根据测试集的预测结果,计算其总体分类准确度、宏平均F1值以及归一化混肴矩阵用于评估模型性能。结果:后融合卷积神经网络具有最高的总体分类准确度(0.855)和宏平均F1值(0.853),根据归一化混淆矩阵,MLC内收误差、外扩误差、无误差、随机误差、平移误差的平均分类准确度分别为0.98、1.00、0.66、0.63、1.00。结论:基于多模态误差图像的后融合卷积神经网络,其在总体分类准确度、宏平均F1值以及每种误差类型的特定分类准确度等方面均显示了该方法的可行性及准确性。Objective To evaluate the feasibility and advantages of multi-channel multi-path DenseNet(MCMP-DenseNet)for detecting multi-leaf collimator(MLC)errors in quality assurance(QA)for intensity-modulated radiotherapy(IMRT)from dose difference maps and Gamma maps.Methods The MLC positions of 98 error-free IMRT plans were modified to simulate translation,extension,shift,and random errors.The plans with and without errors were re-imported into TPS for calculating the dose distributions in the PTW 729 phantom.The dose difference maps and the Gamma maps with two Gamma criteria which were created from the measured and calculated dose distributions were used for dataset establishment and MCMPDenseNet training.Among them,330 dose difference maps and 660 Gamma maps was adopted for the test set,and the remaining were divided into training and validation sets according to 5-fold cross-validation.Based on the prediction results of the test set,the overall classification accuracy,Macro-F1,and normalized confusion matrix were calculated for evaluating the model performance.Results MCMP-DenseNet had the highest overall classification accuracy(0.855)and Macro-F1(0.853).The normalized confusion matrix revealed that the average classification accuracies of the MLC shift error,expansion error,error-free,random error,and translation error were 0.98,1.00,0.66,0.63,and 1.00,respectively.Conclusion The study demonstrates the feasibility and accuracy of MCMP-Densenet in terms of the overall classification accuracy,Macro-F1,and specific classification accuracy.

关 键 词：深度学习 误差分类 调强放射治疗 质量保证 

分 类 号：R318[医药卫生—生物医学工程] R811.1[医药卫生—基础医学]