基于U-net卷积神经网络宫颈癌磁共振临床靶区和危及器官自动勾画的应用  

Application of automatic sketching of clinical target volume and organs at risk in cervical cancer based on U-net convolutional neural network

作　　者：李霞[1] 刘娅 王聪[2] 李振江[2] LI Xia;LIU Ya;WANG Cong;LI Zhenjiang(Heze Municipal Hospital,Shandong Province,Heze274000,China;Shandong Cancer HospitalShandong Institute of Cancer Prevention and Treatment,Shandong Province,Jinan250012,China)

机构地区：[1]山东省菏泽市立医院,山东菏泽274000 [2]山东省肿瘤医院山东省肿瘤防治研究院,山东济南250012

出　　处：《中国医药导报》2022年第24期98-102,共5页China Medical Herald

摘　　要：目的应用U-net卷积神经网络建立基于磁共振影像的宫颈癌临床靶区(CTV)和危及器官(OARs)的自动勾画模型。方法收集2019年4月至2020年12月山东省肿瘤医院大孔径磁共振定位并完成根治性放疗的宫颈癌ⅡB~ⅣA期患者43例。对患者磁共振影像的组织结构信息进行研究,人工勾画感兴趣区(ROI),包括CTV和OARs(膀胱、直肠、左股骨头、右股骨头)。采用计算机进行简单随机抽样,将43例患者分为训练集35例,验证集4例,测试集4例。应用U-net卷积神经网络构建训练模型,验证后对测试集ROI进行自动勾画。比较人工勾画与自动勾画的耗时及Dice相似系数(DSC)值。结果患者平均自动勾画耗时为(44.5±0.6)s,短于平均人工勾画的(2280.0±356.7)s,差异有统计学意义(P<0.05)。自动勾画与人工勾画的DSC值:直肠为(0.752±0.049);CTV为(0.831±0.038);膀胱为(0.943±0.016);左股骨头为(0.894±0.009);右股骨头为(0.896±0.004)。结论U-net卷积神经网络结合磁共振图像可以较为准确地实现CTV和OARs的自动勾画,提高实际临床工作效率。Objective To establish the automatic sketching model of clinical target volume(CTV)and organs at risks(OARs)of cervical cancer based on magnetic resonance images using U-net convolutional neural network.Methods A total of 43 patients with stageⅡB-ⅣA cervical cancer who underwent radical radiotherapy by large aperture magnetic resonance imaging in Shandong Cancer Hospital from April 2019 to December 2020 were enrolled.The tissue structure information of the patient’s magnetic resonance imaging was studied,and the region of interest(ROI)was delineated manually,including CTV and OARs(bladder,rectum,left femoral head,and right femoral head).A total of 43 patients were divided into training set(35 cases),validation set(4 cases)and test set(4 cases)by computer simple random sampling.U-net convolutional neural network was used to construct the training model,and the test set ROI was automatically sketching after verification.The time consumption and Dice similarity coefficient(DSC)value of manual sketching and automatic sketching were compared.Results The average automatic sketching time of patients was(44.5±0.6)s,which was shorter than that of manual sketching(2280.0±356.7)s,and the difference was statistically significant(P<0.05).The DSC value of automatic and manual sketching:the rectum was(0.752±0.049);CTV was(0.831±0.038);bladder was(0.943±0.016);left femoral head was(0.894±0.009);right femoral head was(0.896±0.004).Conclusion U-net convolutional neural network combined with magnetic resonance images can accurately realize the automatic sketching of CTV and OARs,and improve the practical efficiency of clinical work.

关 键 词：磁共振 宫颈癌 自动勾画 U-net 

分 类 号：R737.3[医药卫生—肿瘤]