MDF-Net:一种多尺度细节特征融合的视网膜血管分割算法
MDF-Net:a multi-scale detail feature fusion algorithm for retinal vessel segmentation
蔡鹏飞 1李碧原 1孙高伟 1李士心1
作者信息
- 1. 天津职业技术师范大学电子工程学院,天津 300222
- 折叠
摘要
视网膜血管密集且不规则分布,许多毛细血管融入背景,对比度较低,导致视网膜血管分割非常复杂.基于编码器-解码器的视网膜血管分割网络由于多次编码和解码,会导致细节特征的不可逆损失,进而导致血管分割错误.针对这些问题,提出一种用于视网膜血管分割的多尺度细节特征融合网络(multi-scale detail feature fusion network,MDF-Net).为了确保在精细血管分割过程中有效提取复杂特征,构建细节增强编码器(detail-enhanced encoder,DEE)模块以增强细节表示能力;引入动态解码器(dynamic decoder,DYD)模块,在解码过程中保留空间信息,减少上采样操作引起的信息损失;采用多尺度特征融合(multi-scale feature fusion,MFF)模块来融合编码和解码过程中的特征图,以实现多尺度上下文信息的有效聚合.将MDF-Net算法与其他 9 种算法在DRIVE、CHASEDB1、STARE数据集上进行对比实验.实验结果表明:MDF-Net算法在DRIVE、CHASEDB1和STARE数据集上的灵敏度(sensitivity,Sen)值分别为0.8250、0.880 9 和 0.863 4,曲线下面积(area under the curve,AUC)值分别为 0.988 5、0.990 8 和 0.990 9,MDF-Net在视网膜血管分割方面表现出卓越的性能.
Abstract
The segmentation of retinal vessels is highly challenging because retinal vessels are densely and irregularly dis-tributed with many capillaries blending into the background,exhibiting low contrast.Moreover,the encoder-decoder-based network for retinal vessel segmentation suffers from irreversible loss of detailed features due to multiple encoding and decod-ing operations,leading to incorrect vessel segmentation.To solve these issues,we propose a multi-scale detail feature fu sion network(MDF-Net)for retinal vessel segmentation.This network incorporates a detail-enhanced encoder(DEE)module to enhance the representation of intricate details,ensuring effective retention of complex features during the seg-mentation of fine vessels.The dynamic decoder(DYD)module is introduced to preserve spatial information during the decod-ing process and minimize the information loss caused by upsampling operations.The multi-scale feature fusion(MFF)mod-ule fuses feature maps from both encoding and decoding and achieves effective aggregation of multi-scale contextual infor mation.The experiments on the DRIVE,CHASEDB1,and STARE datasets reveal that MDF-Net achieves sensitivity(Sen)values of 0.825 0,0.880 9,and 0.863 4,as well as area under the curve(AUC)values of 0.988 5,0.9908,and 0.9909,respectively.These results demonstrate that MDF-Net significantly outperforms existing algorithms in retinal vessel segmen-tation.
关键词
视网膜血管分割/U-Net/细节增强/动态上采样/多尺度特征融合Key words
retinal vessel segmentation/U-Net/detail enhancement/dynamic upsampling/multi-scale feature fusion引用本文复制引用
基金项目
天津市自然科学基金资助项目(2021FQ-0027)
天津市教委科研计划项目(2020KJ124)
天津职业技术师范大学科研启动项目(41401-KRKC012007)
出版年
2024
NETL
NSTL
万方数据