基于深度学习方法的放疗患者摆位误差预测  

Prediction of setup errors for patients treated with radiotherapy based on deep learning method

作　　者：高翔[1] 宋双[1] 张伟[1] 陈妙然 夏宇 曹征[1] GAO Xiang;SONG Shuang;ZHANG Wei;CHEN Miaoran;XIA Yu;CAO Zheng(Hematology & Oncology Department, The First People’s Hospital of Hefei, Hefei 230001)

机构地区：[1]合肥市第一人民医院血液肿瘤科,合肥230001

出　　处：《北京生物医学工程》2020年第4期380-388,共9页Beijing Biomedical Engineering

摘　　要：目的为了实现对放疗患者日常摆位误差的准确预测,优化锥形束计算机断层扫描(cone beam computed tomography,CBCT)使用频率,确保患者摆位误差处在允许范围内的同时尽可能减少其承受的额外辐射剂量,本研究基于深度学习方法能够对复杂体系进行预测的能力,构建了一种深度全连接神经网络。方法选取20名头颈部肿瘤患者累积共76次CBCT扫描结果作为研究对象。首先,根据文献调研及临床实践经验,确定患者日常治疗时的摆位误差受到不同技术人员的工作经验、患者体型、固定膜的松紧程度、靶区位置、靶区大小及形状等因素的影响,且与患者前三次治疗时的摆位误差相关性较强,因此针对患者每次CBCT扫描得到的摆位误差,从患者的治疗记录及患者CT图像中获得相关信息,得到467个特征值作为深度学习的输入值,以三个方向上最大摆位误差的分类作为深度学习的输出值,将每次摆位误差的最大值以3 mm为标准分为两类,作为深度学习的目标值。然后,将研究数据按7∶3的比例随机划分为训练集和验证集,通过训练集训练深度神经网络,再使用验证集对神经网络进行初步评估。在完成深度神经网络的训练后,将正在治疗中的新患者的数据作为测试集,使用训练好的神经网络预测新患者的摆位误差大小,并与实际结果对比,评估其准确率。最后,进行重复实验,判断神经网络的预测结果是否具有可重复性。结果本研究构建的深度神经网络对患者摆位误差的预测准确率可以达到86%,能准确预测患者摆位误差大于3 mm的情况,且预测结果的可重复性好。结论基于深度学习方法可以较准确地预测放疗患者日常摆位误差的最大绝对值是否大于3 mm,为优化CBCT扫描频率提供了可靠依据,有助于提高放疗疗效,减轻放疗副反应,具有良好的临床应用价值。Objective By accurately predicting the daily setup errors of each radiotherapy patient,the use of cone beam computed tomography(CBCT)could be optimized to reduce the risks of additional radiation exposure for the patients.For this purpose,a deep fully connected neural network was proposed in this paper.Methods A total of 76 CBCT results from different patients with head and neck tumors were selected for this study and.First,based on literature research and clinical practice experience,it was determined that the setup errors of patients during daily treatment were affected by the work experience of different technicians,the patients’body shape,the tightness of the fixed membranes,the locations,the sizes and shapes of the tumors.And it had a strong correlation with the setup errors during the first three treatments of the patients.Therefore,for each CBCT verification result of the patients,relevant information was obtained from the patients’treatment records and the patients’CT images,and 467 characteristic values were obtained.A set of 467 feature values was used as the input values of deep learning,and the maximum values of the setup errors in three directions were used as the output values.The maximum value of each setup error was divided into two categories with the standard of 3 mm as the label for deep learning.Then,the research data was randomly divided into a training set and a validation set according to the ratio of 7∶3.The deep neural network was trained by the training set,and evaluated by the validation set.After completing the training of the neural network,the data of new patients were used as the test set.The trained neural network and test set were used to predict the daily setup errors of new patients,and the results were compared with the actual results to evaluate its prediction accuracy.Finally,repeated experiments were performed to determine whether the training and prediction results of the neural network were repeatable.Results The prediction accuracy of the patient’s setup errors

关 键 词：深度学习 放疗 摆位误差 锥形束CT 预测 

分 类 号：R318.6[医药卫生—生物医学工程] R815.6[医药卫生—基础医学]