基于深度卷积神经网络的人工智能在喉鳞状细胞癌窄带成像辅助诊断中的应用  

Application of deep convolutional neural networks in the diagnosis of laryngeal squamous cell carcinoma based on narrow band imaging endoscopy

作　　者：胡蓉[1] 钟琦[1] 徐文[1] 黄志刚[1] 程丽宇 王元 何雨蓉 成迎端 Hu Rong;Zhong Qi;Xu Wen;Huang Zhigang;Cheng Liyu;Wang Yuan;He Yurong;Cheng Yingduan(Department of Otorhinolaryngology Head and Neck Surgery,Beijing Tongren Hospital,Capital Medical University,Key Laboratory of Otorhinolaryngology Head and Neck Surgery(Capital Medical University),Ministry of Education,Beijing 100730,China;Department of Urology,the First Affiliated Hospital of Southern University of Science and Technology,the Second Clinical Medical College of Jinan University,Shenzhen People′s Hospital,Shenzhen 518000,China)

机构地区：[1]首都医科大学附属北京同仁医院耳鼻咽喉头颈外科,耳鼻咽喉头颈外科教育部重点实验室,北京100730 [2]南方科技大学第一附属医院,暨南大学第二临床医学院深圳市人民医院泌尿科,深圳518000

出　　处：《中华耳鼻咽喉头颈外科杂志》2021年第5期454-458,共5页Chinese Journal of Otorhinolaryngology Head and Neck Surgery

摘　　要：目的探讨基于卷积神经网络(convolutional neural network,CNN)的人工智能(artificial intelligence,AI)技术通过深度学习辅助喉鳞状细胞癌(以下简称喉鳞癌)临床诊断的可行性。方法本研究采用一套深度CNN用以评估喉鳞癌患者的窄带成像(narrow band imaging,NBI)内镜图像。纳入2015—2017年期间就诊于首都医科大学附属北京同仁医院耳鼻咽喉头颈外科的喉病变患者4799例,其中男3168例,女1631例,年龄21~87岁。采用简单随机化法选取2427例患者的NBI内镜(其中喉良性病变1388例,喉鳞癌1039例)用于对AI系统的训练和校正。对余下的2372例患者采用NBI内镜(其中喉良性病变1276例,喉鳞癌1096例)对AI进行测试,并与耳鼻咽喉头颈外科专家判读结果进行比较。采用SPSS 21.0软件进行卡方检验,计算AI及耳鼻咽喉头颈外科专家判读的准确率、敏感度及特异度,采用受试者工作特征曲线(receiver operating curve,ROC)的曲线下面积(area under the curve,AUC)来评估本算法对NBI内镜图像的判读能力。结果AI验证集的准确率为90.91%(AUC=0.96),敏感度为90.12%,特异度为91.53%,与耳鼻咽喉头颈外科专家判读结果相当[准确率为(91.93±3.20)%,敏感度为(91.33±3.25)%,特异度为(93.02±2.59)%],差异无统计学意义(t值分别为0.64、0.75、1.17,P值分别为0.32、0.28、0.21)。CNN的判读速度明显高于耳鼻咽喉头颈外科专家,差异有统计学意义(每图0.01 s比每图5.50 s,t=9.15,P<0.001)。结论本研究证实了基于深度CNN的AI在喉NBI内镜判读上的有效性,提示AI在喉鳞癌的临床辅助诊断方面有很好的应用前景。Objective To explore the possibility of using artificial intelligence(AI)technology based on convolutional neural network(CNN)to assist the clinical diagnosis of laryngeal squamous cell carcinoma(LSCC)through deep learning algorithm.Methods A deep CNN was developed and applied in narrow band imaging(NBI)endoscopy of 4799 patients with laryngeal lesions,including 3168 males and 1631 females,aged from 21 to 87 years,from 2015 to 2017 in Beijing Tongren Hospital,Capital Medical University.A simple randomization method was used to select the laryngeal NBI images of 2427 patients(1388 benign lesions and 1039 LSCC lesions)for the training and correction the CNN model.The remaining laryngeal NBI images of 2372 patients(including 1276 benign lesions and 1096 LSCC lesions)were used as validation data set to compare performance between CNN and otolaryngologists.SPSS 21.0 software was used for Chi-square test to calculate the accuracy,sensitivity and specificity of AI and otolaryngologists.The area under the curve(AUC)of receiver operating curve(ROC)was used to evaluate the diagnostic ability of the algorithm for NBI images.Results The accuracy,sensitivity and specificity for NBI predictions were respectively 90.91%(AUC=0.96),90.12%and 91.53%,which were equivalent to those for otolaryngologists′predictions(accuracy,sensitivity and specificity were(91.93±3.20)%,(91.33±3.25)%and(93.02±2.59)%,t values were 0.64,0.75 and 1.17,and P values were 0.32,0.28 and 0.21,respectively).The diagnostic efficiency of CNN was significantly higher than that of otolaryngologists(0.01 vs.5.50,t=9.15,P<0.001).Conclusion AI based on deep CNN is effective for using in the laryngeal NBI image diagnosis,showing a good application prospect in the diagnosis of LSCC.

关 键 词：喉肿瘤 喉镜 窄带成像 卷积神经网络 人工智能 

分 类 号：R739.65[医药卫生—肿瘤]