A generative adversarial network (GAN) in the semi-supervised learning architecture was used to address the problem of the scarcity of labeled data in X-ray image classification. Initially, we used a softmax layer to replace the output layer of an unsupervised GAN, extending it to a semi-supervised GAN. In addition, we defined additional labels for the GAN-synthesized samples to guide the training process and optimized the network parameters using a semi-supervised training strategy. Then, the discriminator network obtained by the training was used for X-ray image classification. From tested front-view chest X-ray images of six lung diseases, we find that the proposed method substantially enhances the supervised learning with limited labeled data. Further, the proposed method demonstrates superior classification performance compared with other semi-supervised methods.