多尺度注意力解析网络的视网膜血管分割方法  

Retinal Vessel Segmentation Method Based on Multi-Scale Attention Analytic Network

作　　者：罗文劼 韩国庆 田学东 Luo Wenjie;Han Guoqing;Tian Xuedong(School of Cyber Security and Computer,Hebei University,Baoding,Hebei 071002,China)

机构地区：[1]河北大学网络空间安全与计算机学院,河北保定071002

出　　处：《激光与光电子学进展》2021年第20期431-444,共14页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61472023);河北省自然科学基金(F2019201329)。

摘　　要：视网膜血管分割是检测多种眼病的重要手段,在视网膜疾病自动筛查系统中发挥重要作用。针对现存方法对细小血管分割不足且易出现病理误分割的问题,提出了一种基于多尺度注意力解析网络的分割算法。该网络以编码-解码架构为基础,在子模块中引入注意力残差块,加强了特征传播能力,降低了光照不均、对比度低对模型的影响;在编码器和解码器之间增加跳跃连接并去掉传统池化层,以保留足够的血管细节信息;运用并行多分支结构和空间金字塔池化两种多尺度特征融合方法,实现不同感受野下的特征提取,提升血管分割性能。实验结果表明,该方法在CHASEDB1和STARE标准集上的F1值分别达到了83.26%和82.56%,灵敏度分别达到了83.51%和81.20%,其性能优于当前主流方法。Retinal blood vessel segmentation is an important means to detect a variety of eye diseases,and it plays an important role in automatic screening systems for retinal diseases.Aiming at the problems of insufficient segmentation of small blood vessels and pathological mis-segmentation by existing methods,a segmentation algorithm based on the multi-scale attention analytic network is proposed.The network is based on the encodingdecoding architecture and introduces attention residual blocks in sub-modules,therefore enhancing the feature propagation ability and reducing the effects of uneven illumination and low contrast on the model.The jump connection is added between the encoder and decoder and the traditional pooling layer is removed to retain sufficient blood vessel detail information.Two multi-scale feature fusion methods,parallel multi-branch structure and spatial pyramid pooling,are used to achieve feature extraction under different receptive fields and improve the performance of blood vessel segmentation.Experimental results show that the F;value of this method on the CHASEDB1 and STARE standard sets reaches 83.26% and 82.56%,the sensitivity reaches 83.51% and 81.20%,respectively,and the proposed method is better than that of current mainstream methods.

关 键 词：医用光学 图像处理 血管分割 编码-解码 注意力残差块 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]