Retinal vessel segmentation plays an important role in the automatic assessment of eye health. Deep learning technology has been extensively employed in medical image segmentation. Specifically, U-net based methods achieve great success in medical image segmentation. However, due to its continuous pooling layer and convolution operation, the spatial information and texture information of the image are destroyed. To address this issue, we propose a SAT-Net that integrates side attention and dense atrous convolution block, which also consists of multi-scale input so that the network can retain more features of the image of the encoder stage. The dense atrous convolution block enables multiple receptive field fusion, which preserves the context information of the image, and the side attention mechanism further enhances the high-level information of the encoded features and reduces the noise in the feature map. We apply this method to different retinal image segmentation datasets and compare with the other methods. The experimental results demonstrate the effectiveness of the proposed method.

视网膜血管分割在自动评估眼睛健康中起重要作用。深度学习技术已广泛应用于医学图像分割中。具体而言，基于U-net的方法在医学图像分割中取得了巨大的成功。但是，由于其连续的合并层和卷积操作，图像的空间信息和纹理信息被破坏了。为了解决这个问题，我们提出了一个SAT-Net，它集成了侧面注意力和密集的无规卷积块，它也由多尺度输入组成，因此网络可以保留编码器级图像的更多特征。密集的无规卷积块可实现多个接收场融合，从而保留图像的上下文信息，侧面注意机制进一步增强了编码特征的高级信息，减少了特征图中的噪声。我们将此方法应用于不同的视网膜图像分割数据集，并与其他方法进行比较。实验结果证明了该方法的有效性。