3D U-net应用于鼻咽癌危及器官自动分割的研究  

Application of 3D U-net in automatic segmentation of OARs in nasopharyngeal carcinoma

作　　者：王继平 李鑫 陈传喜 杨志勇 常敦瑞 章桦 王萍[1] WANG Ji-ping;LI Xin;CHEN Chuan-xi;YANG Zhi-yong;CHANG Dun-rui;ZHANG Hua;WANG Ping(Oncology Centre,Huanggang Central Hospital,Yangtze University,Huanggang 438000,Hubei Province,China;Department of Radiation Oncology,the Second People's Hospital of Yichang,Yichang 443000,Hubei Province,China;The Research and Development Center of Beijing Linkingmed Technology Corporation,Beijing 100085,China)

机构地区：[1]长江大学附属黄冈市中心医院肿瘤中心,湖北黄冈438000 [2]宜昌市第二人民医院肿瘤放射治疗科,湖北宜昌443000 [3]北京连心医疗科技有限公司研发中心,北京100085

出　　处：《医疗卫生装备》2020年第11期17-20,45,共5页Chinese Medical Equipment Journal

摘　　要：目的:基于3D U-net提出一种自动分割模型,评估其自动分割效率和几何学精度,探讨其应用于临床的可行性。方法:选择58例鼻咽癌患者的CT定位图像,由1名有经验的放疗医师根据ICRU 83号报告以及中国鼻咽癌协作组调强放疗靶区和危及器官勾画共识手工勾画眼球、晶体、视交叉、视神经、垂体、颞叶、脑干、脊髓、腮腺、内耳、喉、下颌骨、下颌关节等危及器官,并将勾画好的数据传输至连心智能放疗云平台进行训练,深度学习后建立鼻咽癌危及器官自动分割数据库。另选取10例鼻咽癌患者作为测试组,分别采用自动分割和手工勾画方法勾画各危及器官,并对这2种勾画方式所需的时间、危及器官体积差异、重叠比(overlap ratio,OR)、戴斯相似性系数(Dice similarity coefficient,DSC)进行比较。结果:对于10例测试对象,自动分割平均时间较手工勾画平均时间节省了1231 s(74.02%)。除视交叉外,其他器官自动分割较手工勾画均大幅节省时间,差异有统计学意义(P<0.05)。脑干、脊髓、颞叶、喉、腮腺、下颌骨、眼球、下颌关节和内耳等有很高的DSC和OR值(均>0.8),重合性很好;晶体和视神经这类小体积器官平均DSC和OR值也在0.7以上,重合性较好;视交叉DSC平均值只有0.58,OR值仅为0.48,重合性较差。结论:基于3D U-net的自动分割模型对鼻咽癌危及器官进行自动分割,可得到很高的精度,应用于临床工作中可大大提高医师工作效率。Objective To evaluate the efficiency and geometric accuracy of the automatic segmentation based on 3D U-net,and to explore the feasibility of the model in clinical application.Methods CT localization images of 58 patients with nasopharyngeal carcinoma were selected,and the eyes,crystals,optic chiasma,optic nerve,pituitary gland,temporal lobe,brainstem,spinal cord,parotid gland,inner ear,larynx,mandible and mandibular joint were manually outlined by an experienced radiologist according to the International Commission on Radiation Units and Measurements(ICRU)Report No.83 and the consensus of Chinese Nasopharyngeal Carcinoma Collaborative Group on Intensified Radiotherapy Target Area and Endangered Organ delineation.The obtained data were transmitted to an intelligent radiotherapy cloud platform for training,and then a database for automatic segmentation of OARs in nasopharyngeal carcinoma was established after deep learning.In addition,10 nasopharyngeal carcinoma patients were selected into a test group,and each organ at risk was sketched by automatic segmentation and manual sketching,and the two methods were compared over the time required,the difference in the volume of the organ at risk,the overlap ratio(OR)and Dice similarity coefficient(DSC).Results For the 10 test subjects,the automatic segmentation had the mean time being 1231 s(74.02%),which was shorter than that by manual sketching.Except the optic chiasma,all the other organs had the sketching time significantly reduced by the automatic segmentation when compared with manual method(P<0.05).There were high DSC and OR values(all>0.8)in brainstem,spinal cord,temporal lobe,larynx,parotid gland,mandible,eyeball,mandibular joint and inner ear.The average DSC and OR values of small organs such as lens and optic nerve were above 0.7 with high OR;the optic chiasma had the mean values of DSC and OR being 0.58 and 0.48 respectively with low OR.Conclusions The automatic segmentation model based on 3D U-net can automatically segment the organs at risk of nasopharyngea

关 键 词：3D U-net 自动分割 人工智能 鼻咽癌 危及器官 重叠比 相似性系数 

分 类 号：R739.63[医药卫生—肿瘤] R730.44[医药卫生—临床医学]