改进D-Linknet的眼底视网膜血管分割

Improved fundus retinal vascular segmentation in D-Linknet

临床医生可通过观察眼底视网膜血管及其分支对人体是否患有疾病进行早期诊断,但由于视网膜中的血管错综复杂,模型在分割时会出现对微细血管分割精确度不足的问题。为此,提出一种结合残差模块Res2-net以及高效通道注意力机制(efficient channel attention,ECA)的D-Linknet模型。首先,利用Res2-net代替基础模型中的残差模块Res-net以提升每个网络层的感受野;其次,在Res2-net中添加一种结合压缩激励(squeeze and excitation,SE)和门通道(gated channel transformation,GCT)的注意力机制模块,改善处于复杂背景下的血管分割效果和效率;在网络的解码层加入ECA确保模型计算的性能,避免因降维导致的精度下降;最后,融合改进的模型输出图与掩膜图细化分割结果。在公开数据集DRIVE、STARE上进行分割实验,模型准确度(accuracy,AC)分别为97.11%、96.32%,灵敏度(sensitivity,SE)为84.55%、83.92%,曲线下方范围的面积(area under curve,AUC)为0.9873和0.9766,分割效果优于其他模型。实验证明了算法的可行性,为后续研究提供科学依据。

Clinicians can make early diagnosis of diseases in the human body by observing retinal blood vessels and their branches in the fundus.However,due to complexity of blood vessels in the retina,the model may be insufficient in accuracy in the segmentation of microvessels.To this end,a D-Linknet model is proposed,which combines residual module(Res2-net)and efficient channel attention(ECA)mechanism.Firstly,Res2-net is used to replace the residual module(Res net)in the basic model to enhance the receptive field of each network layer;Secondly,an attention mechanism module combining squeeze and excitation(SE)and gated channel transformation(GCT)is added to Res2-net to improve the effectiveness and efficiency of vascular segmentation under complex backgrounds;ECA is added to the decoding layer of the network to ensure the performance of model calculation,avoiding accuracy degradation caused by dimensionality reduction;Finally,the improved model output image and mask image are fused to refine the segmentation results.Segmentation experiments have been conducted on public datasets DRIVE and STARE,with model accuracy(AC)of 97.11%and 96.32%,sensitivity(SE)of 84.55%and 83.92%,and area under curve(AUC)of 0.9873 and 0.9766,respectively.The segmentation performance is superior to other models,demonstrating feasibility of the algorithm.This paper provides a scientific basis for subsequent research.