计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (8): 166-170.DOI: 10.3778/j.issn.1002-8331.1910-0097

• 图形图像处理 • 上一篇    下一篇

基于注意力机制的乳腺X线摄影分类方法

盛龙帅,李策,李欣   

  1. 1.江苏省社会安全图像与视频理解重点实验室(南京理工大学),南京 210094
    2.中国矿业大学(北京) 机电与信息工程学院,北京 100083
  • 出版日期:2020-04-15 发布日期:2020-04-14

Classification of Mammography Based on Attention Mechanism

SHENG Longshuai, LI Ce, LI Xin   

  1. 1.Jiangsu Key Laboratory of Image and Video Understanding for Social Safety, Nanjing University of Science and Technology, Nanjing 210094, China
    2.School of Mechanical Electronic and Information Engineering, China University of Mining and Technology(Beijing), Beijing 100083, China
  • Online:2020-04-15 Published:2020-04-14

摘要:

环境的日益恶化导致癌症的发病率不断升高,2018年全球乳腺癌的发病率在所有癌症中已经位居首位。乳腺X线摄影价格实惠且易于操作,目前被认作是最好的乳腺癌筛查方法,也是早期发现乳腺癌最有效的方法。针对乳腺X线摄影不容易分辨、特征不明显等特点,提出了基于RNN+CNN的注意力记忆网络对其进行分类。注意力记忆网络包含注意力记忆模块和卷积残差模块。注意力记忆模块中,注意力模块提取乳腺X线摄影的特征,记忆模块在RNN网络加入注意力权重来模拟人对所提取关键信息的重点突出;卷积残差模块使用CNN对图像进行分类。该方法创新之处在于:提出注意力记忆网络用于乳腺X线摄影图像分类;所设计网络在RNN+CNN结构上引入注意力权重,提取图像关键信息以增强特征描述。在乳腺X线摄影INbreast数据集上的实验结果显示,注意力记忆网络的运行时间比预训练的Inceptionv2、ResNet50、VGG16网络少50%以上,同时达到更高的分类准确率。

关键词: 乳腺X线摄影, 注意力机制, 卷积神经网络, 循环神经网络, 预训练

Abstract:

The worsening environment has led to an increase in the incidence of cancer. In 2018, the incidence of breast cancer has ranked first among all cancers in the world. Considering that mammography is affordable and easy to operate, it is currently regarded as the best screening method for breast cancer and the most effective method for early detection of breast cancer. In view of the fact that mammography is not easy to distinguish and its features are not obvious, this paper proposes an attention memory network based on RNN + CNN to classify it. The attention memory network includes the attention memory module and the convolution residual module. In the attention memory module, the attention module is used to extract features of mammography, and the memory module adds attention weight to RNN network to simulate people’s emphasis on key information. Convolution residual module uses CNN to classify images. The contributions of this paper are as follows, the attention memory network is proposed for mammography image classification, the designed network introduces attention weights on the RNN+CNN structure, and extracts key information of images to enhance the feature description. The experimental results on the mammography INbreast dataset show that the runtime of attention memory network can be 50% less than the pre-trained Inception v2, ResNet50, and VGG16, and can achieve higher classification accuracy than others.

Key words: mammography, attention mechanism, Convolutional Neural Network(CNN), Recurrent Neural Network(RNN), pre-training