Abstract:

As one of the most common cancers for women, the breast cancer deadly threatens patients’s health, which make multiple classification of mammorgraphy image an important role in clinical diagnosis of breast cancer. Multiple classification research of classical Convolutional Neural Network（CNN） on the breast mammorgraphy image directly takes use of its advanced features. Such method is not accurate enough. While the construction of human network model increases the accuracy of classification. It functions with extracting low-level features of images from stacked convolutional layer and maximum pooling layer, and gradually passing them back to images. By extracting high-level features to cascade the lower one, features of cascading appear, which will be finally classified after being pooled through the global maximum pooling layer. In accordance with a simulation experiment on 1,824 breast mammorgraphy images by Cancer Center, Sun Yat-sen University, the construction of human network model rises accuracy rate to 74.54%, that make it way better than relevant net work models.

Key words: Convolutional Neural Network（CNN）, mammography image, low-level feature, advanced feature, feature fusion

摘要：

乳腺癌是女性最常见的恶性肿瘤之一，严重威胁患者健康，因此乳腺钼靶图像多分类对临床诊断乳腺癌具有十分重要的作用。传统卷积神经网络直接采用高级特征对乳腺钼靶图像进行多分类研究，此方法准确率不高。为了进一步提高分类准确率，构建人型网络模型进行分类。此结构通过堆叠的卷积层以及最大池化层来进行图片的低级特征进行提取，通过堆叠的卷积层以及上池化层将特征逐步返回到图片形式的特征图，通过堆叠的卷积层以及最大池化层再次提取到更高级的特征并与之前的低级特征进行级联，将级联的特征经过全局最大池化层进行池化并得到最终分类。在中山大学肿瘤防治中心的1 824幅乳腺钼靶图像做仿真实验，实验结果表明，该方法的准确率达到了74.54%，优于现有相关网络模型。

关键词: 卷积神经网络, 乳腺钼靶图像, 低级特征, 高级特征, 特征融合