基于Swin Transformer和卷积注意力的乳腺癌病理图像诊断研究  

Research on breast cancer pathological image diagnosis based on Swin Transformer and convolutional attention

作　　者：禤浚波[1,2,3] 周立广 梁英豪 梁淑慧 付志鸿 关志广 毕明霞 XUAN Junbo;ZHOU Liguang;LIANG Yinghao;LIANG Shuhui;FU Zhihong;GUAN Zhiguang;BI Mingxia(Guangxi Key Lab of Multi-source Information Mining&Security,Guangxi Normal University,Guilin 541004,China;School of Artificial Intelligence,Nanning Vocational and Technical University,Nanning 530008,China;School of Computer Science and Engineering,Guangxi Normal University,Guilin 541004,China;Wuxiang Hospital of the Second Nanning People′s Hospital,Nanning 530219,China;Qingdao Sino-German Intelligent Manufacturing Technician College,Qingdao 266555,China)

机构地区：[1]广西师范大学广西多源信息挖掘与安全重点实验室,广西桂林541004 [2]南宁职业技术大学人工智能学院,广西南宁530008 [3]广西师范大学计算机科学与工程学院,广西桂林541004 [4]南宁市第二人民医院五象医院,广西南宁530219 [5]青岛中德智能制造技师学院,山东青岛266555

出　　处：《现代电子技术》2024年第13期36-42,共7页Modern Electronics Technique

基　　金：广西自然科学基金项目(2022GXNSFAA035625);广西多源信息挖掘与安全重点实验室开放基金项目(MIMS21-02);山东省职业教育教师创新团队(物联网专业)资助;青岛市信息技术名师工作室资助。

摘　　要：为了降低由于医生阅片疲劳或经验不足而可能导致的漏诊或误诊问题,提高医生诊断乳腺癌病理图像的准确性和工作效率,文中采用北京大学国际医院提供的公开的最大乳腺癌病理组织图像数据集,包括正常、良性病变、原位癌和浸润癌四种类型,并提出了一种基于Swin Transformer和卷积注意力机制的乳腺癌病理图像诊断方法,给出了诊断算法的框架和处理流程,在评价指标方面取得了96.93%的精确率、97.82%的召回率和97.74%的准确率,与常用的卷积神经网络ResNet152、VGG16相比,精确率和准确率都是最高的,从而证明提出的方法是有效的。最后,基于Flask技术和Swin Transformer开发了可视化的乳腺癌病理图像诊断软件,只需提供一张患者的乳腺癌病理组织图像,10 s左右即可自动输出诊断结果,可以极大地提高医生的工作效率。In order to reduce the missed diagnosis and misdiagnosis caused by fatigue or inexperience of doctors,and to improve the accuracy and work efficiency of doctors in diagnosing breast cancer,the public largest data set of breast cancer pathological images provided by Peking University International Hospital,including four types,named normal,benign,carcinoma in situ and invasive carcinoma,is adopted,and a method of breast cancer pathological image diagnosis based on Swin Transformer and convolutional attention mechanism is proposed,and the frame and processing flow of the diagnosis algorithm are given.In terms of the evaluation indexes,the precision ratio of the proposed method is 96.93%,its recall rate is 97.82%and its accuracy rate is 97.74%.In comparison with the common convolutional neural network(CNN)ResNet152 and VGG16,its precision rate and accuracy rate are the highest.Therefore,it is proved that the proposed method is effective.Finally,on the basis of the technology of Flask and Swin Transformer,a visual diagnosis software of breast cancer pathological image is developed,which can automatically output the diagnosis results in about 10 seconds only by providing a patient′s breast cancer pathological image,which can greatly improve the work efficiency of doctors.

关 键 词：乳腺癌 病理图像 深度学习 Swin Transformer 卷积注意力机制 FLASK 

分 类 号：TN911.73-34[电子电信—通信与信息系统]