基于深度学习超声在乳腺肿块四分类中的应用价值  

Application value of deep learning ultrasound in the four-category classification of breast masses

作　　者：于腾飞[1] 何文[1] 甘从贵 赵明昌 张红霞[1] 宁彬[1] 宋海曼[1] 郑帅 李怡 朱宏远 Yu Tengfei;He Wen;Gan Conggui;Zhao Mingchang;Zhang Hongxia;Ning Bin;Song Haiman;Zheng Shuai;Li Yi;Zhu Hongyuan(Department of Ultrasonography,Beijing Tiantan Hospital,Capital Medical University,Beijing 100070,China;CHISON Medical Technologies Co.,LTD,Wuxi 214142,China;Institute for Infocomm Research,Agency for Science and Technology Research,138632 Singapore)

机构地区：[1]首都医科大学附属北京天坛医院超声科,100070 [2]无锡祥生医疗科技股份有限公司开发部,214142 [3]新加坡科研局信息通讯研究院,新加坡138632

出　　处：《中华超声影像学杂志》2020年第4期337-342,共6页Chinese Journal of Ultrasonography

基　　金：国家重点研发计划(2016YFC0104801);中国超声医师科技新星计划(2018)。

摘　　要：目的:探讨采用基于卷积神经网络(CNN)构建的人工智能辅助诊断模型对乳腺肿块四分类进行超声鉴别诊断的应用价值。方法:利用CNN构建辅助诊断模型,收集2016年1月至2018年1月首都医科大学附属北京天坛医院的2098例乳腺肿块患者(其中良性肿瘤1132例,恶性肿瘤779例,炎症32例,腺病155例)的10490张超声图像,将其分为训练集和测试集,对人工智能辅助诊断模型进行训练及测试。并将分别使用二维成像(two-dimensional imaging,2D)和二维彩色多普勒成像(two dimensional color Doppler flow imaging,2D-CDFI)的两组数据训练模型进行比较。分析乳腺良性肿瘤、恶性肿瘤、炎症、腺病患者的ROC曲线,计算ROC曲线下面积(AUC)。结果:使用2D-CDFI数据训练的模型比使用2D数据训练的模型对良性肿瘤和炎症的训练集和验证集的准确性有显著提高。①对于良性肿瘤的诊断,使用2D图片训练集的敏感性92%,特异性95%,AUC 0.93;使用2D-CDFI图片训练集的敏感性93%,特异性95%,AUC 0.93;使用2D图片测试集的敏感性91%,特异性96%,AUC 0.94;使用2D-CDFI图片测试集的敏感性93%,特异性94%,AUC 0.94。②对于恶性肿瘤的诊断,使用2D图片训练集的敏感性93%,特异性97%,AUC 0.94;使用2D-CDFI图片训练集的敏感性93%,特异性96%,AUC 0.94;使用2D图片测试集的敏感性93%,特异性96%,AUC 0.94;使用2D-CDFI图片测试集的敏感性93%,特异性96%,AUC 0.94。③对于炎症的诊断,使用2D图片训练集的敏感性81%,特异性99%,AUC 0.91;使用2D-CDFI图片训练集的敏感性86%,特异性99%,AUC 0.89;使用2D图片测试集的敏感性100%,特异性98%,AUC 0.98;使用2D-CDFI图片测试集的敏感性100%,特异性99%,AUC 0.96。④对于腺病的诊断,使用2D图片训练集的敏感性88%,特异性97%,AUC 0.94;使用2D-CDFI图片训练集的敏感性93%,特异性98%,AUC 0.94;使用2D图片测试集的敏感性94%,特异性98%,AUC 0.93;使用2D-CDFI图片测试集的敏感性88%,特异性99%Objective To explore the application value of artificial intelligence-assisted diagnosis model based on convolutional neural network(CNN)in the differential diagnosis of benign and malignant breast masses.Methods A total of 10490 images of 2098 patients with breast lumps(including 1132 cases of benign tumor,779 cases of malignant tumor,32 cases of inflammation,155 cases of adenosis)were collected from January 2016 to January 2018 in Beijing Tiantan Hospital Affiliated to the Capital University of Medical Sciences.They were divided into training set and test set and the auxiliary artificial intelligence diagnosis model was used for training and testing.Two sets of data training models were compared by two-dimensional imaging(2D)and two-dimensional and color Doppler flow imaging(2D-CDFI).The ROC curves of benign breast tumors,malignant tumors,inflammation and adenopathy were analyzed,and the area under the ROC curve(AUC)were calculated.Results The accuracies of 2D-CDFI ultrasonic model for training group and testing group were significantly improved.①For benign tumors,the result from training set with 2D image was:sensitivity 92%,specificity 95%,AUC 0.93;the result from training set with 2D-CDFI images was:sensitivity 93%,specificity 95%,AUC 0.93;the result for test set with 2D images was:sensitivity 91%,specificity 96%,AUC 0.94;the result for test set with 2D-CDFI images was:sensitivity 93%,specificity:94%,AUC 0.94.②For malignancies,the result for training set with 2D images was:sensitivity 93%,specificity 97%,AUC 0.94;the result for training set with 2D-CDFI images was:sensitivity 93%,specificity 96%,AUC 0.94;the result for test set with 2D images was:sensitivity 93%,specificity 96%,AUC 0.94;the result for test set with 2D-CDFI images was:sensitivity 93%,specificity 96%,AUC 0.94.③For inflammation,the result for training set with 2D images was:sensitivity 81%,specificity 99%,AUC 0.91;the result for training set with 2D-CDFI images was:sensitivity 86%,specificity 99%,AUC 0.89;the result for test set with 2D i

关 键 词：卷积神经网络 人工智能 辅助诊断 乳腺肿块 

分 类 号：R737[医药卫生—肿瘤] R44[医药卫生—临床医学]