结合注意力机制和多路径U-Net的视网膜血管分割

Retinal Vessel Segmentation Based on Attention Mechanism and Multi-Path U-Net

针对现有算法无法精确分割细微血管末端,且分割结果易受光学造影与病变区域影响的问题,提出一种结合注意力和多路径U-Net的视网膜血管分割算法.首先,设计一个双路径U-Net,通过纹理与结构分支提取粗和细粒度血管,并使用语义指导模块充分融合深浅层特征;其次,采用一种引入注意力机制和DropBlock的残差模块来代替普通卷积模块,改善处于复杂背景区域中血管的分割效果,防止过拟合;最后,将双路径U-Net的输出图与原图传入特征细化模块进行特征提取和融合,进一步细化血管分割结果.在DRIVE,STARE和CHASEDB1数据集上的实验结果表明,该算法的准确率分别为97.01%,96.43%和97.52%;灵敏度分别为80.31%,84.38%和81.61%;受试者工作特性曲线下方的面积(AUC)分别为98.67%,98.06%和98.83%,综合分割性能优于其他算法.

To solve the problem that the existing algorithms cannot accurately segment the end of tiny blood vessels and the segmentation results are easily affected by optical contrast and the diseased area,a retinal vessel segmentation algorithm is proposed,which combining attention mechanism and multi-path U-Net.Firstly,a dual-path U-Net is designed to extract coarse and fine-grained blood vessels separately through texture and structure branches.The semantic guidance module is designed to fully integrate the deep and shallow features.Secondly,a residual block introducing attention mechanism and DropBlock is designed to replace the ordinary convolution block,which improves the segmentation effect of blood vessels in complex background area and prevents over-fitting.Finally,the output image of the dual-path U-Net and the original image are put into the feature refinement module for feature extraction and fusion,which further refines the blood vessel segmentation results.The accuracy of the proposed algorithm on DRIVE,STARE and CHASEDB1 datasets are 97.01%,96.43%,and 97.52%;the sensitivity are 80.31%,84.38%and 81.61%;and the AUC(Area Under Curve)are 98.67%,98.06%,and 98.83%,respectively.The experimental results show that the comprehensive segmentation performance of the proposed algorithm is better than other algorithms.