基于深度学习的鼻咽癌放疗临床靶区体积和危及器官自动勾画研究  

Automatic contouring of clinical target volume and organs at risk in radiotherapy for nasopharyngeal carcinoma based on deep learning

作　　者：钱杰伟 陈雪梅[1] 李军[1] 程品晶[2] 单国平[3] 张获 桂龙刚 柏正璐 Qian Jiewei;Chen Xuemei;Li Jun;Cheng Pinjing;Shan Guoping;Zhang Huo;Gui Longgang;Bai Zhenglu(Oncology Department,Northern Jiangsu People's Hospital,Yangzhou 225001,China;School of Nuclear Science and Technology,University of South China,Hengyang 421001,China;Key Laboratory of Radiation Oncology,Zhejiang Cancer Hospital,Hangzhou 310022,China)

机构地区：[1]江苏省苏北人民医院肿瘤科,江苏扬州225001 [2]南华大学核科学技术学院,湖南衡阳421001 [3]浙江省肿瘤医院放射肿瘤学重点实验室,浙江杭州310022

出　　处：《实用肿瘤杂志》2024年第6期536-541,共6页Journal of Practical Oncology

基　　金：浙江省放射肿瘤学重点实验室开放课题(2022ZJCCRAD05);江苏省苏北人民医院科研基金项目(yzucms202004)。

摘　　要：目的构建一种基于深度学习U-Net的鼻咽癌临床靶区体积(clinical target volume,CTV)和危及器官(organs at risk,OARs)的自动勾画方法,并与基于图谱的自动勾画方法(atlas-based auto-segmentation,ABAS)比较,进而探讨基于深度学习自动勾画方法的可行性与优越性。方法选取2022年1月至9月于江苏省苏北人民医院行鼻咽癌放疗的患者的CT定位影像150例并进行预处理,构建基于U-Net的自动勾画模型,其中90例作为训练数据集,10例作为验证集,其余50例作为测试集,以医师手工勾画结果为金标准,计算U-Net自动勾画模型对鼻咽癌CTV和OARs的自动勾画精度,并与ABAS勾画结果进行比较。结果U-Net的CTV和OARs(脑干、脊髓、左眼球、右眼球、左晶状体、右晶状体、左视神经、右视神经、左下颌骨、右下颌骨、左腮腺、右腮腺、左颞叶和右颞叶)的戴斯相似性系数(Dice similarity coefficient,DSC)分别为(0.76±0.03)、(0.93±0.02)、(0.92±0.03)、(0.93±0.02)、(0.94±0.03)、(0.90±0.03)、(0.91±0.02)、(0.78±0.06)、(0.77±0.05)、(0.95±0.04)、(0.95±0.02)、(0.80±0.04)、(0.81±0.03)、(0.77±0.05)和(0.76±0.04)。除CTV、视神经、腮腺和颞叶外,U-Net模型自动勾画的其余器官的豪斯多夫距离(Hausdorff distance,HD)值均≤5.60 mm且重叠比(overlap ratio,OR)值均≥0.80。U-Net较ABAS模型自动勾画的各个器官的DSC更高,HD更低且OR更高(均P<0.05),勾画各个器官的耗时也更少,总体耗时降低(176.73±54.08)s(P<0.05)。结论U-Net自动勾画模型能较好实现鼻咽癌放疗CTV和OARs的自动勾画,为临床医师的勾画提供参考并提高勾画效率,以深度学习为基础的自动勾画方法具有很高的可行性和优越性。Objective To construct a U-Net model based on deep learning to realize the automatic contouring of clinical target volume(CTV)and organs at risk(OARs)of radiotherapy plan for nasopharyngeal carcinoma(NPC)patients,and analyze its feasibility and superiority as compared with atlas-based auto-segmentation(ABAS)method.Methods The CT images of 150 NPC patients undergoing radiotherapy at Northern Jiangsu People's Hospital,from January to September 2022,were selected and preprocessed to construct an automatic segmentation model based on U-Net.Ninety cases were used as a training set,10 cases as a verification set and the remaining 50 cases as a test set.The contouring accuracy of the U-Net automatic contouring model for the CTV and OARs of NPC were calculated and compared with the automatic contouring module of ABAS.Results The Dice similarity coefficients(DSCs)of CTV and OARs including brainstem,spinal cord,left eye,right eye,left lens,right lens,left optic nerve,right optic nerve,left mandible,right mandible,left parotid gland,right parotid gland,left temporal lobe,and right temporal lobe,were(0.76±0.03),(0.93±0.02),(0.92±0.03),(0.93±0.02),(0.94±0.03),(0.90±0.03),(0.91±0.02),(0.78±0.06),(0.77±0.05),(0.95±0.04),(0.95±0.02),(0.80±0.04),(0.81±0.03),(0.77±0.05),and(0.76±0.04),respectively.Except for CTV,optic nerves,parotid glands and temporal lobes,the Hausdorff distances(HDs)of other organs were≤5.60 mm and the overlap ratios(ORs)were≥0.80.Compared to ABAS,U-Net had higher DSCs,lower HDs,and higher ORs for the automatic contouring of CTV and OARs(all P<0.05).U-Net also took less time to delineate each organ,and reduced the overall time consumption by(176.73±54.08)seconds(P<0.05).Conclusions U-Net realized the automatic contouring of CTV and OARs in NPC radiotherapy,and improved the contouring efficiency for clinicians.The automatic contouring model based on deep learning had high feasibility and advantages.

关 键 词：鼻咽癌 深度学习 U-Net 放射治疗 自动勾画 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] R739.63[医药卫生—肿瘤]