文献类型：期刊文献

中文题名：基于卷积神经网络的乳腺病理图像识别算法

英文题名：Recognition algorithm of breast pathological images based on convolutional neural network

作者：凌语[1];孙自强[1]

机构：[1]华东理工大学信息科学与工程学院

年份：2019

卷号：40

期号：5

起止页码：573

中文期刊名：江苏大学学报（自然科学版）

外文期刊名：Journal of Jiangsu University：Natural Science Edition

收录：CSTPCD;;Scopus;北大核心:【北大核心2017】；

基金：中央高校基本科研业务费专项资金资助项目(222201917006)

语种：中文

中文关键词：乳腺肿瘤;卷积神经网络;VGG-19A;BN算法;dropconnect算法;迁移学习

外文关键词：breast tumor;convolutional neural network;VGG-19A network;BN algorithm;dropconnect algorithm;transfer learning

摘要：为了协助病理医生诊断乳腺肿瘤,提出了一种计算机自动识别分析乳腺病理图像的方法.该方法采用乳腺病理图像数据集BreaKHis作为数据样本,在卷积神经网络模型VGG-19的基础上,提出了一种VGG-19A的改进网络模型.通过在卷积层中的激活函数前加入BN算法,在全连接层中加入dropconnect层,来优化网络模型的性能,提升网络模型的识别精度.此外,考虑到迁移学习方法能够让网络模型更加充分地学习图像特征,将其引入到VGG-19A网络的训练中.将该网络应用到乳腺病理图像的识别过程中,同时采用PFTAS+QDA,PFTAS+SVM,PFTAS+RF,Single-TaskCNN,AlexNet以及VGG-19算法进行了对照试验.结果表明新算法在图像识别的准确性和泛化性能上相较现有方法都有了一定的提升,因而具有一定的临床应用价值.
To assist pathologists in diagnosing breast tumors, a method was proposed to automatically recognize and analyze breast pathological images by computers. With BreaKHis as data sample, the improved network model of VGG-19A was proposed based on the convolutional neural network model of VGG-19. By adding BN algorithm before activation function in convolution layer and adding dropconnect layer in full connection layer, the performance of network model was optimized to improve the recognition accuracy of network model. Considering that the transfer learning method could make the network model learn pathological features more fully, the method was introduced into the training of VGG-19A network. The network was applied to the recognition of breast pathological images, and PFTAS+QDA, PFTAS+SVM, PFTAS+RF, Single-Task CNN, AlexNet and VGG-19 algorithms were used in comparative experiments. The results show that compared with the existing methods, the proposed method can improve the accuracy and generalization performance of image recognition, which has important practical application value.