DMDR-UNet:一种眼底视网膜血管分割算法  

DMDR-UNet:An algorithm for retinal blood vessel segmentation

作　　者：孙君顶[1] 张宏英 SUN Junding;ZHANG Hongying(School of Computer Science and Technology,Henan Polytechnic University,Jiaozuo 454000,Henan,China)

机构地区：[1]河南理工大学计算机学院,河南焦作454000

出　　处：《河南理工大学学报（自然科学版）》2023年第6期142-148,共7页Journal of Henan Polytechnic University（Natural Science）

基　　金：国家自然科学基金资助项目(62276092);河南省科技攻关项目(212102310084);河南省高等学校重点科研项目(22A520027)。

摘　　要：针对眼底视网膜血管存在微血管特征采集困难的问题,基于U-Net模型,提出一种改进的眼底视网膜血管分割算法DMDR-UNet(deformable multiscale dense residual-UNet)。首先,根据视网膜血管走势自由、形态丰富的特点,提出可变形卷积网络模块,增强微血管特征提取;其次,基于多尺度空洞卷积模块,聚合不同感受野下视网膜血管的多尺度信息;最后,设计密集残差连接模块,减少编解码间语义鸿沟的同时加强特征信息的交互与补充。基于DRIVE数据集进行实验,结果表明,本文方法能够准确识别并分割出视网膜微血管,分割效果更好。Aiming at the difficulty of collecting microvascular features of retinal vessels in fundus,an im‐proved DMDR-UNet(deformable multiscale dense residual-UNet)segmentation model was proposed based on U-Net model.Firstly,a deformable convolutional network module was proposed to enhance the extraction of microvascular features,considering the free trend and rich morphology of retinal vessels.Secondly,multi-scale cavity convolution module was used to aggregate multi-scale information of retinal vessels under differ‐ent receptive fields.Finally,a dense residual connection module was designed to reduce the semantic gap be‐tween codecs and enhance the interaction and supplement of feature information.Experiments based on DRIVE dataset showed that the proposed method could accurately identify and segment retinal microves‐sels,and achieve better segmentation results.

关 键 词：视网膜血管分割 U-Net 可变形卷积 多尺度 密集残差 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]