In the field of image recognition, images with insufficient label data cannot be well recognized by the supervised method model. In order to solve this problem, a semi-supervised method model based on Generative Adversarial Network(GAN) was proposed. That is, by combining the advantages of semi-supervised GANs and deep convolutional GANs, and replacing the sigmoid activation function with softmax in the output layer, the Semi-Supervised Deep Convolutional GAN(SS-DCGAN) model was established. Firstly, the generated samples were defined as pseudo-samples and used to guide the training process. Secondly, the semi-supervised training method was adopted to update the parameters of the model. Finally, the recognition of abnormal(cerebral infarction) images was realized. Experimental results show that the SS-DCGAN model can recognize abnormal images well with little label data, which achieves 95.05% recognition rates. Compared with Residual Network 32(ResNet32) and Ladder networks, the SS-DCGAN model has significant advantages.