MCFNet:融合上下文信息的多尺度视网膜动静脉分类网络

MCFNet:Multi-scale retinal artery and vein classification network fused with context information

针对由于血管类间具有强相似性造成的动静脉错误分类问题,提出了一种新的融合上下文信息的多尺度视网膜动静脉分类网络(multi-scale retinal artery and vein classification network,MCFNet),该网络使用多尺度特征(multi-scale feature,MSF)提取模块及高效的全局上下文信息融合(efficient global contextual information aggregation,EGCA)模块结合U型分割网络进行动静脉分类,抑制了倾向于背景的特征并增强了血管的边缘、交点和末端特征,解决了段内动静脉错误分类问题。此外,在U型网络的解码器部分加入3层深度监督,使浅层信息得到充分训练,避免梯度消失,优化训练过程。在2个公开的眼底图像数据集(DRIVE-AV,LES-AV)上,与3种现有网络进行方法对比,该模型的F1评分分别提高了2.86、1.92、0.81个百分点,灵敏度分别提高了4.27、2.43、1.21个百分点,结果表明所提出的模型能够很好地解决动静脉分类错误的问题。

Aiming at the problem of wrong classification of arteries and veins due to the strong similarity between blood vessels,this paper proposes a new multi-scale retinal artery and vein classification network(MCFNet)that integrates context information.In combination with the U-shaped segmentation network,the network uses multi-scale feature(MSF)extraction module and high-efficiency global contextual information aggregation(EGCA)module to classify the arteries and veins,which suppresses the features that tend to the background,enhances the edge,intersection and end features of the blood vessels,solving the problem of wrong classification of arteries and veins in the segment.In addition,the decoder part of the U-shaped network is added with three layers of depth supervision to fully train the shallow information,so as to avoid the disappearance of gradient and optimize the training process.On two open retina image data sets(DRIVE-AV,LES-AV),compared with three existing network methods,the F1 score of this model has increased by 2.86,1.92,and 0.81 percentage points,respectively,and the sensitivity increased by 4.27,2.43,and 1.21 percentage points,respectively.The results show that the proposed model can well solve the problem of error in classification of arteries and veins.