基于多模态超声图像的深度学习在鉴别诊断乳腺良恶性肿块中的应用  

Application of deep learning based on multimodal ultrasound images in differential diagnosis of benign and malignant breast masses

作　　者：李金瑶 柳懿垚 姜伟 Li Jinyao;Liu Yiyao;Jiang Wei(Ultrasound Department,Union Shenzhen Hospital,Huazhong University of Science and Technology,Shenzhen 518060,China;School of Biomedical Engineering,Health Science Center,Shenzhen University,Shenzhen 518060,China)

机构地区：[1]华中科技大学协和深圳医院超声科,广东深圳518060 [2]深圳大学医学部生物医学工程学院,广东深圳518060

出　　处：《实用肿瘤杂志》2025年第1期20-27,共8页Journal of Practical Oncology

基　　金：深圳市科创委基金面上项目(JCYJ20220530142002005)。

摘　　要：目的探讨基于单独B型(B-mode)、应变弹性成像(strain elastography,SE)、自动乳腺全容积成像(automated breast volume scanner,ABVS)、联合B型和SE(B-mode+SE)、联合B型和ABVS(B-mode+ABVS)、联合SE和ABVS(SE+ABVS)以及联合B型、SE和ABVS(B-mode+SE+ABVS)图像所构建的不同深度学习(deep learning,DL)模型对辅助诊断乳腺良恶性肿块的效能。方法回顾性分析2021年8月至2023年8月在华中科技大学协和深圳医院超声科进行乳腺肿块超声检查的病例。对所纳入的病例的超声图像进行数据预处理,包括肿块范围、感兴趣区分割和数据增强。将图像输入多模态交互融合模型训练,运用DL的方法,构建鉴别乳腺良恶性肿块的7种DL模型,分别是B-mode-DL、SE-DL、ABVS-DL、B-mode+SE-DL、B-mode+ABVS-DL、SE+ABVS-DL和B-mode+SE+ABVS的多模态DL(Mutimodal-DL)模型。采用受试者工作特征(receiver operating characteristic,ROC)曲线的曲线下面积(area under the curve,AUC)对DL模型的诊断效能进行评价。结果纳入508例病例,采用简单随机抽样方法分为训练集284例(良性250例,恶性34例)和测试集224例(良性199例,恶性25例)。7种DL模型中,Multimodal-DL模型诊断乳腺良恶性肿块的准确度、敏感度和特异度均最高,分别为95.4%、95.2%和95.5%,与其他模型比较差异均具有统计学意义(均P<0.05);AUC和Youden指数也最高,分别为0.955和0.907。结论基于多模态超声图像所建立的DL模型对乳腺良恶性肿块的鉴别效果最好。Objective To explore the performance of different deep learning(DL)models based on the images of either B-mode ultrasound,strain elastography(SE),automated breast volume scanner(ABVS),B-mode ultrasound combined with SE(B-mode+SE),B-mode ultrasound combined with ABVS(B-mode+ABVS),SE combined with ABVS(SE+ABVS),or B-mode ultrasound combined with SE and ABVS(B-mode+SE+ABVS)in assisting the differential diagnosis of benign and malignant breast masses.Methods A retrospective analysis was performed on the cases of breast mass ultrasonography from August 2021 to August 2023 in the ultrasound department of Union Shenzhen Hospital of Huazhong University of Science and Technology.Data preprocessing was performed on the ultrasound images of the cases,including mass extent,the segmentation of regions of interest,and data enhancement.The images were fed into the multimodal interactive fusion model for training,and DL method was used to construct seven DL models to distinguish benign and malignant breast masses.The seven DL models were based on either B-mode,SE,ABVS,B-mode+SE,B-mode+ABVS,SE+ABVS,or B-mode+SE+ABVS(multimodal)images.Area under the receiver operating characteristic curve was used to evaluate the diagnostic performance of the models.Results A total of 508 cases were included and divided into a training set of 284 cases,including 250 benign cases and 34 malignant cases,and a test set of 224 cases including 199 benign cases and 25 malignant cases,by simple random sampling.The accuracy,sensitivity,and specificity of the multimodal-DL model were 95.4%,95.2%,and 95.5%,respectively,which were higher than those of the other six DL models(all P<0.05).The AUC and Youden index of the multimodal-DL model were 0.955 and 0.907,which were greater than those of the other six DL models.Conclusions The DL model based on multimodal ultrasound images has the best performance on the differentiation of benign and malignant breast masses.

关 键 词：乳腺肿瘤 人工智能 多模态超声 自动乳腺全容积成像 深度学习 

分 类 号：R73[医药卫生—肿瘤]