图像分割是医学图像分析中的一个基本问题，基于深度学习的典型UNet架构（UNet architecture）分割网络模型及其变式被广泛应用于视网膜血管分割之中。但是UNet网络通过局部卷积模块提取图像的特征信息，难以关联图像的全局信息，无法有效捕捉像素之间的长距离依赖关系。考虑到UNet网络模型存在的问题和视网膜血管图像的特点，在UNet跳跃连接中加入注意力模块，可以捕捉血管之间的长距离依赖关系。此外，为增强网络的分割能力，使用群归一化（group normalization, GN）代替UNet网络模型原始的批归一化（batch normalization, BN），对不同的通道选择对应的分组。为更新参数和优化网络，利用侧输出层和最后的输出层设计了交叉熵损失函数。在DRIVE数据集和CHASEDB1数据集上进行了实验，实验结果表明所提出的模型有更好的图像分割效果。

Image segmentation is a fundamental problem in medical image analysis, the typical deep learning based UNet architecture (UNet) and its variants are widely used in retinal vessel segmentation. However, the UNet network extracts feature information from images through local convolution modules, which makes the global information of the images difficult to be correlated and the long-distance dependencies between pixels difficult to be effectively captured. Considering the problems with the UNet network model and the characteristics of retinal vascular images, an attention module was added to the skip connections of UNet to capture long-distance dependencies between blood vessels. In addition, to enhance the segmentation ability of the network, the group normalization(GN) was used instead of the original batch normalization (BN) of the UNet network model, and the corresponding groups were selected for different channels. To update parameters and optimize the network, the final cross entropy loss function was designed using the side output layer and the final output layer. Experiments are implemented on the DRIVE dataset and CHASEDB1 dataset, and the experimental results show that the proposed model has better image segmentation performance.