基于双分支多视角深度自注意力编码器的乳腺肿瘤分类方法  

Classification of Breast Tumors Based on Dual-Branch Multi-View Transformer

作　　者：柳懿垚 杨意 陈敏思 汪天富 姜伟 雷柏英 Liu Yiyao;Yang Yi;Chen Minsi;Wang Tianfu;Jiang Wei;Lei Baiying(Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging,Shenzhen 518060,Guangdong,China;Department of Biomedical Engineering,School of Medicine,Shenzhen University,Shenzhen 518060,Guangdong,China;Department of Ultrasonics,Huazhong University of Science and Technology Union Shenzhen Hospital,Shenzhen 518052,Guangdong,China)

机构地区：[1]广东省生物医学信息检测和超声成像重点实验室,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 [3]华中科技大学协和深圳医院超声医学科,广东深圳518052

出　　处：《中国生物医学工程学报》2022年第5期527-536,共10页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(62001302)

摘　　要：自动乳腺全容积超声成像(ABVS)系统因其高效、无辐射等特性成为筛查乳腺癌的重要方式。针对ABVS图像进行计算机辅助乳腺肿瘤良恶性分类的研究,有利于帮助临床医生准确、快速地进行乳腺癌的诊断,甚至可辅助提高低年资医生的诊断水平。ABVS系统产生的三维乳腺图像数据量较大,造成常规的深度学习方式训练时间长、占用资源巨大。本研究设计了一种基于ABVS数据的多视角图像提取方式,替代常规的三维数据输入,在降低参数量的同时弥补二维深度学习中的空间关联性;其次,基于交叉视角图像的空间位置关系,提出一种深度自注意力编码器(Transformer)网络,用于获得图像的有效特征表达。实验是基于自有ABVS数据库的153例容积图像,良恶性分类的准确率为86.88%,F1-评分为81.70%,AUC达到0.8316。所提出的方法有望应用于ABVS图像的乳腺肿瘤良恶性筛查。Automatic breast volume scanner(ABVS)system is the primary screening method for breast cancer because of its high efficiency and no radiation.The study of computer-aided breast cancer classification based on ABVS images is helpful for clinicians to diagnose breast cancer accurately and quickly and can even help to improve the diagnostic level of junior doctors.Because of its imaging mode,ABVS system produces a large amount of three-dimensional breast image data,leading to a long training time and huge resources of conventional deep learning.Therefore,we designed a multi-view image extraction method based on ABVS data,which replaced the conventional 3 D data and made up for the spatial correlation in 2 D deep learning while reducing the number of parameters.Secondly,based on the spatial position relationship of cross view images,we proposed a self-attention encoder(Transformer)to obtain effective feature expression of the images.Our experiment was based on 153 volume images from our own ABVS database.The accuracy of benign and malignant classification was 86.88%,the F1 score was 81.70%and AUC reached 0.8316.The experimental results indicated that the proposed method could be effectively applied to the benign and malignant screening of breast tumors based on ABVS images.

关 键 词：乳腺癌 多视角 自注意力编码器 良恶性诊断 自动乳腺全容积超声成像 

分 类 号：R737.9[医药卫生—肿瘤] TP391.41[医药卫生—临床医学]