Skin lesion is regarded as an illness, which is often detected and diagnosed by utilizing dermoscopy imaging. The essential step of skin lesion analysis is to segment the lesion region accurately. However, many variations in view and scale of lesion areas make segmentation quite challenging. In this paper, we propose an effective skin lesion segmentation architecture based on a modified generative adversarial network. As the network can learn great feature representation of dermoscopic images automatically in an unsupervised way, supervised learning could make it more promising. In particular, we explore an encoder-decoder network concatenating residual blocks and atrous convolution as the generator to generate the segmentation masks. This generator can obtain deep representation and preserve fine-grained information. The decision of the discriminator is very important to optimize a network. To enhance the reliability of the discrimination, two discriminators are applied to distinguish generated masks from ground truth jointly. The model is trained and evaluated using ISIC skin lesion challenge datasets in 2017. Our experiments demonstrate that the proposed network achieves the favorable segmentation performance compared with the other CNN-based supervised methods.