深度卷积神经网络在头颈部肿瘤小体积危及器官自动勾画中的应用  

Application of automatic small volume organ at risk segmentation based on deep convolution neural network in 250 cases of head and neck tumors

作　　者：康盛伟 吴骏翔 唐斌[1,2] 杨凤 侯氢 KANG Sheng-wei;WU Jun-xiang;TANG Bin;YANG Feng;HOU Qing(Key Laboratory of Radiation Physics and Technology,Ministry of Education,Institute of Nuclear Science and Technology,Sichuan University,Chengdu 610064,China;Sichuan Cancer Hospital&Institute,Sichuan Cancer Center,Cancer Hospital Affiliate to School of Medicine,University of Electronic Science and Technology of China,Radiation Oncology Key Laboratory of Sichuan Province,Chengdu 610041,China)

机构地区：[1]四川大学原子核科学技术研究所,辐射物理及技术教育部重点实验室,四川成都610064 [2]四川省肿瘤医院·研究所,四川省癌症防治中心,电子科技大学医学院附属肿瘤医院,放射肿瘤学四川省重点实验室,四川成都610041

出　　处：《中华肿瘤防治杂志》2022年第8期571-577,共7页Chinese Journal of Cancer Prevention and Treatment

基　　金：四川省重点研发项目(2021YFG0168,2021YFG0320)。

摘　　要：目的评估基于分类模型的三维(3D)编码-解码深度卷积神经网络在自动勾画头颈部危及器官中的几何学精度和分割效率,为临床使用提供依据.方法选取2020-01-05-2020-11-21四川省肿瘤医院收治的250例头颈部肿瘤患者CT,由资深放射治疗医师手动勾画危及器官,以其中150套图像作为训练集,用于调整自动勾画网络参数,60套为验证集,40套作为测试集.使用CT分类模型将患者CT在头脚方向分成6部分,通过3 D编码-解码分割网络自动勾画脑干、脊髓、腮腺、下颌骨、气管、食管、垂体、眼球、晶体、视神经、视交叉和耳蜗等12个危及器官.采用戴斯相似性系数和豪斯多夫距离评估自动勾画的精度.采用单因素方差分析及配对t检验等方法对数据进行统计学分析.结果训练集所有危及器官自动勾画的戴斯相似性系数平均值为0.94,豪斯多夫距离平均值为2.22 mm;验证集的戴斯相似性系数平均值为0.82,豪斯多夫距离平均值为2.96 mm;测试集戴斯相似性系数平均值为0.84,均>0.70,83.33%的危及器官戴斯相似性系数>0.80,豪斯多夫距离平均值为2.87 mm,91.67%的危及器官豪斯多夫距离<3.50 mm.自动勾画网络勾画效率较高,平均勾画时间为125 s.结论基于分类模型的3D编码-解码分割网络自动勾画头颈部肿瘤的危及器官可以得到较准确的结果,尤其是小体积危及器官的勾画,为头颈部肿瘤放射治疗危及器官自动勾画提供了方法.Objective To evaluate the geometric accuracy and segmentation efficiency of a classification model-based three-dimensional(3 D) encoding-decoding deep convolutional neural network in automatically delineating organs at risk in the head and neck, and to provide evidence for clinical use.Methods The CT of 250 patients with head and neck tumors admitted to Sichuan Cancer Hospital from January 1 to February 21,2020 was selected, and the organ at risk was manually delineated by a senior radiotherapy physician, and 150 sets of images were used as a training set for automatic adjustment.Outline the network parameters, 60 sets were the validation set, and 40 sets were the test set.Using the CT classification model, the patient CT was divided into 6 parts in the direction of the head and feet, and the brainstem, spinal cord, parotid gland, mandible, trachea, esophagus, pituitary, eyeball, lens, optic nerve, optic chiasm and cochlea were automatically delineated through a 3 D encoding-decoding segmentation network.The Dice similarity coefficient and Hausdorff distance were used to evaluate the accuracy of automatic delineation.One-way ANOVA and paired t-test were used for statistical analysis of data.Results The average dice similarity coefficient of all organs at risk for automatic delineation in the training set was 0.94,and the average Hausdorff distance was 2.22mm.The average dice similarity coefficient of the validation set was 0.82,and the average Hausdorff distance was 2.96mm.The average test set dice similarity coefficient was 0.84,all>0.70,83.33% organs at risk had dice similarity coefficient>0.80.The mean value of Hausdorff distance was 2.87mm,and 91.67% organ at risk Hausdorff distance was less than 3.50mm.The automatic delineation network used in this study had higher delineation efficiency,with an average delineation time of 125s.Conclusion The classification modelbased 3Dencoding-decoding segmentation network can automatically delineate the organs at risk of head and neck tumors,which can obtain more accurate

关 键 词：深度学习 自动勾画 头颈部肿瘤 危及器官 放射治疗 

分 类 号：R730.55[医药卫生—肿瘤] R739.91[医药卫生—临床医学]