基于跨模态注意力机制特征B型超声与弹性超声融合模块联合诊断乳腺良、恶性肿瘤  

B-mode ultrasound combined with elastic ultrasound based on feature fusion module of cross-modal attention mechanism for diagnosis of benign and malignant breast tumors

作　　者：王彤 苏畅[1,2] 何萍 王心怡[3,4] 崔立刚[3] 林伟军[1,2] WANG Tong;SU Chang;HE Ping;WANG Xinyi;CUI Ligang;LIN Weijun(Ultrasonic Laboratory,Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China;Department of Ultrasound Diagnosis,Peking University Third Hospital,Beijing 100191,China;Center of Breast,Beijing Cancer Hospital,Beijing Institute for Cancer Research,Key Laboratory of Carcinogenesis and Transformation Research,Beijing 100142,China)

机构地区：[1]中国科学院声学研究所超声学实验室,北京100190 [2]中国科学院大学电子电气与通讯工程学院,北京100049 [3]北京大学第三医院超声诊断科,北京100191 [4]北京大学肿瘤医院北京市肿瘤防治研究所乳腺中心恶性肿瘤发病机制及转化研究教育部重点实验室,北京100142

出　　处：《中国医学影像技术》2022年第12期1862-1866,共5页Chinese Journal of Medical Imaging Technology

摘　　要：目的设计跨模态注意力机制特征融合模块,观察其用于B型超声与弹性超声联合诊断乳腺良、恶性肿瘤的价值。方法收集371例接受常规超声检查及超声弹性成像的女性乳腺肿瘤患者、共466处病灶;按3∶1∶1将466组病灶图像分为训练集(n=280)、验证集(n=93)及测试集(n=93)。采用卷积神经网络分支模型分别提取B型超声图像和弹性超声图像特征,之后以基于跨模态注意力机制的多模态特征融合网络进行特征融合,观察其诊断乳腺良、恶性肿瘤的价值。结果改进后的DenseNet用于B型超声诊断乳腺良、恶性肿瘤的准确率为88.43%,敏感度为88.96%,特异度为87.31%,其效能略优于改进前。基于跨模态注意机制特征融合的B型超声与弹性超声联合诊断乳腺良、恶性肿瘤的准确率为94.23%,敏感度为95.11%,特异度为93.28%,效能优于决策加权融合模型、直接串联融合模型及单模态模型。结论跨模态注意力机制特征融合模块可在一定程度上提高B型超声与弹性超声联合诊断乳腺良、恶性肿瘤的效能。Objective To design the feature fusion module of cross-modal attention mechanism,and to observe its value for B-mode ultrasound combined with elastic ultrasound in diagnosis of benign and malignant breast tumors.Methods A total of 371 female patients(466 lesions)with breast tumors who underwent conventional ultrasound and ultrasound elastography were enrolled,and the lesion images were divided into training set(n=280),verification set(n=93)and test set(n=93)at the ratio of 3∶1∶1.The features of B-mode ultrasound images and elastic ultrasound images were extracted with convolutional neural network branch model,then the feature fusion was performed with multi-mode feature fusion network based on cross modal attention mechanism,and its value for diagnosing benign and malignant breast tumors was observed.Results The efficacy of improved DenseNet for B-mode ultrasound diagnosis of benign and malignant breast tumors was slightly better than that before improvement,with accuracy of 88.43%,sensitivity of 88.96%and specificity of 87.31%.The efficacy of B-mode ultrasound combined with elastic ultrasound based on feature fusion module of cross-modal attention mechanism for diagnosing benign and malignant breast tumors was better than decision weighted fusion model,direct serial fusion model and single-mode model,with accuracy of 94.23%,sensitivity of 95.11%and specificity of 93.28%.Conclusion Feature fusion module of cross-modal attention mechanism could improve the value of B-mode ultrasound combined with elastic ultrasound for diagnosis of benign and malignant breast tumors in a certain extent.

关 键 词：乳腺肿瘤 神经网络 计算机 超声检查 弹性成像技术 

分 类 号：R737.9[医药卫生—肿瘤] R445.1[医药卫生—临床医学]