融合动态损失与渐进多尺度注意力的视网膜血管分割网络

Progressive Multiscale Attention Network with Dynamic Loss for Retinal Vessel Segmentation

视网膜血管分割任务仍然存在许多挑战,例如眼底图像中生物学背景复杂、毛细血管细小且模糊以及特征难以得到充分利用。针对这些问题,提出了一种融合动态损失与渐进多尺度注意力的网络。其中渐进式特征提取策略可以在网络中逐步捕获对分割有益的特征,并且保留更多细节。设计的多尺度通道注意力模块能够获得复杂的通道依赖关系并且抑制跳跃连接过程中的背景噪音,达到突出重要特征的目的。最后提出动态损失用来自适应调整深监督中各损失函数的权重,优化训练策略。所提出的方法在两个公开的数据集DRIVE和CHASE_DB1中进行了充分的验证,其中灵敏度分别达到0.838 9和0.846 8,准确率分别达到0.971 5和0.974 5,展现出了较好的分割性能。

There are still many challenges in retinal vessel segmentation,such as complex biological backgrounds,small and blurry capillaries,and difficulty in fully utilizing features in fundus images.In response to these issues,this paper proposes a network that integrates dynamic loss and progressive multi-scale attention.The progressive feature extraction strategy can gradually capture features that are beneficial for segmentation in the network,while retaining more details.The designed multi-scale channel attention module can obtain complex channel dependencies and suppress background noise during skip connections,achieving the goal of highlighting important features.Finally,it is proposed that the dynamic loss uses the weights of each loss function in the deep supervision through adaptive adjustment to optimize the training strategy.The proposed method has been fully validated in two publicly available datasets,DRIVE and CHASEDB1,with sensitivity reaching 0.8389 and 0.8468,and accuracy reaching 0.9715 and 0.9745,respectively,demonstrating good segmentation performance.