融合双残差密集与注意力机制的视网膜血管分割

Retinal vascular segmentation based on fusion of double residual density and attention mechanism

针对视网膜血管末端细小,且容易与背景混淆等现象从而导致细小血管不易分割和断裂等情况,提出了一种融合双残差密集与注意力机制的视网膜血管分割算法。首先,在编码器部分利用双残差密集块与高效通道注意力机制来获取特征;其次,为了解决细小血管分割不足的现象,在编码器与解码器中间使用空洞卷积替换标准卷积来增大感受野;最后,自适应聚合块将之前所有块的特征映射组合起来,形成一个新的特征映射,作为后续层的输入,在自适应聚合块或DDRB之后,将使用卷积层来压缩特征映射,则双残差密集块(从DDRB1到DDRB5)的输出特征映射被完全重用。分别在DRIVE和STARE数据集上进行验证,其ACC分别为96.85%和97.84%,AUC分别为98.61%和99.45%。

In response to the problems that retinal blood vessels end are small and easy to be confused with the background,so that small blood vessels are not easy to be divided and broken,the retinal vessel segmentation algorithm combining dual residual density and attention mechanism is proposed.Firstly,in the encoder section,double residual dense blocks and efficient channel attention mechanism are utilized to obtain features.Secondly,in order to solve the problem of insufficient segmentation of small blood vessels,cavity convolution is used between the encoder and the decoder to replace the standard convolution to increase the receptive field.Finally,the adaptive aggregation block combines the feature maps of all previous blocks to form a new feature map as input to subsequent layers.After the adaptive aggregation block or DDRB,convolutional layers will be used to compress the feature map,and the output feature maps of double residual dense blocks(from DDRB1 to DDRB5)will be fully reused.Verified on the DRIVE and STARE datasets,the ACC values are 96.85% and 97.84%,respectively,and the AUC values are 98.61%and 99.45%,respectively.