This paper presents the autoencoder-generative adversarial network (AGAN) in the analysis of mammograms. The AGAN architecture is used to augment the data by generating additional representations of the mammogram images, enhancing this way the information of the analyzed problem. The images generated by this this deep network are appended to the original set of mammograms and fed to the input of convolutional neural network, which plays the role of the final classifier. The proposed system was used to recognize the mammograms belonging to two classes: normal and abnormal. The investigations were performed using a large database consisting of 11218 regions of interest of mammographic images from the DDSM base. The results demonstrate the advantage of this proposed deep learning system over other known approaches to mammogram recognition. Our average accuracy in detecting abnormal cases (malignant plus benign versus healthy) was 89.71%, sensitivity 93.54%, specificity 80.58% and AUC=0.9410. These results are among the best for this large database.

本文介绍了自动编码器生成对抗网络（AGAN）在乳房X线照片的分析中。AGAN体系结构用于通过生成乳房X射线照片图像的其他表示来增强数据，从而以这种方式增强所分析问题的信息。由这个深层网络生成的图像被附加到原始的乳房X线照片上，并被馈送到卷积神经网络的输入，卷积神经网络起着最终分类器的作用。所提出的系统用于识别属于两类的乳房X线照片：正常和异常。研究是使用大型数据库进行的，该数据库包含来自DDSM基地的11218个乳房X线照片的感兴趣区域。结果证明了该提议的深度学习系统相对于其他已知的乳房X线照片识别方法的优势。我们检测异常病例（恶性加良性与健康）的平均准确度为89.71％，敏感性93.54％，特异性80.58％和AUC = 0.9410。对于大型数据库而言，这些结果是最好的。