Abstract: The differences in quality, shooting conditions and acquisition status of different datasets lead to uneven results in model image segmentation. To address this situation, the fundus image optic disc and cup segmentation algorithm under adversarial learning is proposed to segment optic disc (OD) and optic cup (OC) from different fundus image datasets. It improved the generator network based on generating adversarial network by adding densely connected blocks, so that the network could obtain better performance with lower computational cost and shorter training time, and improve the generalization ability of the model in different datasets. Experimental results show that the stability of the algorithm in terms of segmentation performance is verified in REFUGE dataset, and the extension of the algorithm to test fundus datasets from different devices without further training achieves better results.