Retinal vessel segmentation is of great significance for assisting doctors in diagnosis of ophthalmological diseases such as diabetic retinopathy, macular degeneration and glaucoma. This article proposes a new retinal vessel segmentation algorithm using generative adversarial learning with a large receptive field. A generative network maps an input retinal fundus image to a realistic vessel image while a discriminative network differentiates between images drawn from the database and the generative network. Firstly, the proposed generative network combines shallow features with the upsampled deep features to assemble a more precise vessel image. Secondly, the residual module in the proposed generative and discriminative networks can effectively help deep nets easy to optimize. Moreover, the dilated convolutions in the proposed generative network effectively enlarge the receptive field without increasing the amount of computations. A number of experiments are conducted on two publicly available datasets (DRIVE and STARE) achieving the segmentation accuracy rates of 95.63% and 96.84%, and the average areas under the ROC curve of 98.12% and 98.53%. Performance results show that the proposed automatic retinal vessel segmentation algorithm outperforms state‐of‐the‐art algorithms in many validation metrics. The proposed algorithm can not only detect small tiny blood vessels but also capture large‐scale high‐level semantic vessel features.

中文翻译：

---

使用具有大接收域的生成对抗网络进行高质量的视网膜血管分割

视网膜血管分割对协助医生诊断糖尿病性视网膜病变，黄斑变性和青光眼等眼科疾病具有重要意义。本文提出了一种新的视网膜血管分割算法，该算法利用具有较大接收域的生成对抗性学习。生成网络将输入的视网膜眼底图像映射到真实的血管图像，而判别网络将从数据库中提取的图像与生成网络区分开。首先，提出的生成网络将浅层特征与上采样的深层特征相结合，以组装更精确的血管图像。其次，提出的生成网络和判别网络中的残差模块可以有效地帮助易于优化的深层网络。此外，所提出的生成网络中的扩张卷积在不增加计算量的情况下有效扩大了接收场。在两个公开可用的数据集（DRIVE和STARE）上进行了许多实验，分别达到了95.63％和96.84％的分割准确率，ROC曲线下的平均面积为98.12％和98.53％。性能结果表明，在许多验证指标中，提出的自动视网膜血管分割算法都优于最新算法。所提出的算法不仅可以检测细小的血管，而且可以捕获大规模的高级语义血管特征。在两个公开可用的数据集（DRIVE和STARE）上进行了许多实验，分别达到了95.63％和96.84％的分割准确率，ROC曲线下的平均面积为98.12％和98.53％。性能结果表明，本文提出的自动视网膜血管分割算法在许多验证指标上均优于最新算法。所提出的算法不仅可以检测微小的细小血管，而且可以捕获大规模的高级语义血管特征。在两个公开可用的数据集（DRIVE和STARE）上进行了许多实验，分别达到了95.63％和96.84％的分割准确率，ROC曲线下的平均面积为98.12％和98.53％。性能结果表明，本文提出的自动视网膜血管分割算法在许多验证指标上均优于最新算法。所提出的算法不仅可以检测微小的细小血管，而且可以捕获大规模的高级语义血管特征。