基于U型卷积网络的视网膜血管分割方法.pdf

第 43 卷 第 2 期 光 学 仪 器 Vol. 43，No. 2
2021 年 4 月
classical encoder-decoder architecture. The encoder module uses the pre-trained residual module to
fully extract the features of each layer, and the decoder module carries out the up-sampling layer by
layer through the transposed convolution, which increases the multiplex ability of features. The
network introduces the atrous spatial pyramid pooling (ASPP) block in the middle layer to extract
retinal vessel characteristics of different scales. In order to keep consistency in the prediction within
the class, the channel attention block is introduced in the skip connection layer to carry out the
收稿日期 ：2020-07-10
基金项目 ：上海市人工智能专项（2019-RGZN-01077）
作者简介 ：秦晓飞(1982—)，男，高级工程师，研究方向为人工智能算法。E-mail：xiaofei.******@
万方数据第 2 期 秦晓飞，等：基于U型卷积网络的视网膜血管分割方法 • 25 •
adaptive refinement of the features, and the features of different levels are fused. Experimental
results on the DRIVE data set show that compared with the performance of other related
algorithms, this algorithm has the highest sensitivity, specificity and accuracy, the best model
generalization ability, and greatly improves the accuracy of retinal vessel seg