In order to solve the problem of too small dataset when training convolutional neural networks, this paper uses a small dataset training GAN(Generative Adversarial Networks) to generate new samples for data augmentation. The expanded data set is applied to the training image classification model and has a good effect. For the dichotomy problem of Herlev datasets, the original training set was used to train Generative Adversarial Networks, generating a large number of high-quality high-resolution cell images, and expanding each training set up to 24,000 cases. Then the expanded training set is used for classification network training. The accuracy of the validation set in Resnet migration learning is 97% which is higher than the accuracy of dataset of affine transformation expansion, which is 93%, showing that our approach is very effective for image data augmentation. This method can also be used for image data augmentation in other fields.