多级自适应尺度的U型视网膜血管分割算法  

Multi-level adaptive scale U-shaped retinal blood vessel segmentation algorithm

作　　者：梁礼明 詹涛 雷坤 冯骏 谭卢敏[2] Liang Liming;Zhan Tao;Lei Kun;Feng Jun;Tan Lumin(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China;School of Applied Sciences,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,赣州341000 [2]江西理工大学应用科学学院,赣州341000

出　　处：《电子测量技术》2022年第13期130-140,共11页Electronic Measurement Technology

基　　金：国家自然科学基金(51365017,61463018);江西省自然科学基金面上项目(20192BAB205084);江西省教育厅科学技术研究重点项目(GJJ170491)资助。

摘　　要：针对视网膜血管细小和尺度变化复杂的特点,提出一种多级自适应尺度的U型视网膜血管分割算法。首先以编码-解码结构为基础引入残差模块,加强通道特征传播能力。其次在网络底部嵌入多尺度特征提取模块,旨在调整感受野有效地提取多尺度特征。同时在跳跃连接部分增加改进的自适应特征融合模块,促进相邻层次特征之间的有效融合,以提取更多的细小血管特征。最后在解码部分设置侧输出的多级注意结构对多层次特征进行自适应细化。实验结果表明,该算法在DRIVE、STARE和CHASEDB1数据集上准确率分别达到0.9645、0.9694和0.9671,灵敏度分别达到0.8417、0.8465和0.8545,AUC分别达到0.9866、0.9908和0.9877,整体性优于现有算法。Aiming at the characteristics of small retinal vessels and complex scale changes,a multi-level adaptive scale U-shaped retinal vessel segmentation algorithm is proposed.Firstly,the residual module is introduced based on the encoder-decoder structure to enhance the channel feature propagation capability.Secondly,a multi-scale feature extraction module is embedded at the bottom of the network to adjust the receptive field to effectively extract multi-scale features.At the same time,an improved adaptive feature fusion module is added to the skip connection part to promote effective fusion between adjacent hierarchical features to extract more small blood vessel features.Finally,the multi-level attention structure output on the setting side of the decoding part performs adaptive refinement on the multi-level features.The experimental results show that the accuracy of the algorithm on the DRIVE,STARE and CHASEDB1 datasets reaches 0.9645,0.9694 and 0.9671,respectively,the sensitivity reaches 0.8417,0.8465 and 0.8545,and the AUC reaches 0.9866,0.9908 and 0.9877,respectively,and the overall performance is better than the existing algorithms.

关 键 词：视网膜血管分割 U型网络 残差模块 多尺度特征提取 自适应特征融合 多级注意 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]