摘要： 准确的视盘和视杯分割能够有效地辅助青光眼的诊断和监测，从而进一步提高治疗效果。然而，现有方法没有考虑到眼底图像不同通道之间的差异，并且难以实现对视杯边界的精确分割。针对这个问题，提出了一种基于对抗学习和引导机制的网络框架(ALG-Net)，旨在提高视盘和视杯的分割性能。ALG-Net由分割网络和鉴别器两部分组成，在分割网络中构建了引导融合模块，该模块将单通道特征信息与RGB图像特征融合，使网络充分学习眼底图像不同通道之间的差异信息，引导分割网络聚焦于关键区域。ALG-Net网络框架还采用了鉴别器，通过对抗性学习的方式促进分割网络生成更真实的分割结果。在REFUGE和Drishti-GS数据集上进行了广泛的实验评估，实验结果表明，ALG-Net在RUFUGE数据集上视盘和视杯分割的平衡精度分别达到了98.6%和95.9%，在Drishti-GS数据集上也表现出更优异的性能。此外，ALG-Net的分割结果还应用于青光眼分类任务，在RUFUGE数据集上ROC曲线下面积取得了98.3%的效果，相较于经典UNet算法提高了1.5%，为青光眼的早期诊断和监测提供了有力的支持。

Abstract: Accurate segmentation of the optic disc and optic cup can effectively assist in the diagnosis and monitoring of glaucoma, thereby further improving treatment outcomes. However, existing methods do not take into ac-count the differences between different channels of fundus images, and it is difficult to achieve accurate segmentation of the optic cup boundary. To address this problem, a network framework based on adversarial learning and guidance mechanism (ALG-Net) is proposed, aiming to improve the segmentation performance of optic disc and optic cup. ALG-Net consists of two parts: segmentation network and discriminator. A guidance fusion module is constructed in the segmentation network, which fuses the single-channel feature information with the RGB image features, so that the network fully learns the difference information between different channels of the fundus image, and guides the segmentation network to focus on the key regions. The ALG-Net network framework also employs a discriminator to motivate the segmentation network to generate more realistic segmentation results by means of adversarial learning. Extensive experimental evaluations were conducted on the REFUGE and Drishti-GS datasets, and the experimental results showed that ALG-Net achieved a balance accuracy of 98.6% and 95.9% for disc and cup segmentation on the RUFUGE dataset, respectively. It also demonstrated better performance on the Drishti-GS dataset. In addition, the segmentation results of ALG-Net were also applied to glaucoma classification tasks, achieving an area under the ROC curve of 98.3% on the RUFUGE dataset, which was 1.5% higher than the classic UNet algorithm, providing strong support for early diagnosis and monitoring of glaucoma.