In this paper,aiming at the robustness and accuracy of classifying a large number of unlabeled samples with only a small number of labeled samples,we propose an improved semi-supervised generative adversarial networks (SGAN) method for breast cancer image classification.This method uses Softmax function instead of Sigmoid function to realize multi-classification in the output layer.Firstly,the random vector is input into the generation network to generate pseudo samples and be labeled as pseudo sample class for training.Then the real labeled samples,real unlabeled samples and pseudo samples are input into the discrimination network and output as different kinds of probability values.Then the semi-supervised training method is used to update the parameters by back propagation.Finally,the classification of breast cancer pathological images is realized.The number of labeled samples is 25,50,100 and 200 respectively.The final accuracy rate is 95.5%.The experimental results show that the accuracy rate of this algorithm has good robustness when the labeled samples are limited.Compared with the classification methods such as convolution neural networks and transfer learning (TL),the accuracy of this algorithm is significantly improved.