This paper presents a new nested U-shape attention network (NUA-Net) with improved robustness of lesions for effective vascular segmentation in retinal imaging. Unlike most of the current deep learning approaches which rely on vanilla upsample module to recover distinguishable features for segmentation, our attention-based multi-scale network extends the U-shape segmentation network by introducing a novel multi-scale upsample attention (MSUA) module to enhance vessel features in a hierarchical structure. The new approach connects encoder-decoder branches through a nested skip-connection pyramid architecture to extract discriminating retinal features from the rich local details. Experimental evaluations on five publicly available databases DRIVE, STARE, CHASE_DB, IOSTAR and HRF show the NUA-Net achieves 0.8043–0.8511 (Sensitivity), 0.9741–0.99 (Specificity) and 0.9646–0.9794 (Accuracy) respectively. The benchmark by cross-testing and separate-testing presents a state-of-the-art performance and better vessel preservation compared with other approaches.

本文提出了一种新的嵌套 U 形注意网络 (NUA-Net)，它具有改进的病变鲁棒性，可用于视网膜成像中的有效血管分割。与目前大多数依赖于普通上采样模块来恢复可区分特征以进行分割的深度学习方法不同，我们基于注意力的多尺度网络通过引入一种新颖的多尺度上采样注意 (MSUA) 模块来扩展 U 形分割网络在层次结构中增强血管特征。新方法通过嵌套的跳过连接金字塔架构连接编码器-解码器分支，以从丰富的局部细节中提取有区别的视网膜特征。对五个公开可用的数据库 DRIVE、STARE、CHASE_DB、IOSTAR 和 HRF 的实验评估显示 NUA-Net 达到 0.8043–0.8511（灵敏度）、0.9741–0。分别为 99（特异性）和 0.9646–0.9794（准确度）。与其他方法相比，交叉测试和单独测试的基准表现出最先进的性能和更好的血管保存。