基于改进U-Net的视网膜血管分割方法研究  

Research on Retinal Blood Vessel Segmentation Based on Improved U-Net

作　　者：郭峰[1] 黄文博[1] 燕杨[1] GUO Feng;HUANG Wen-bo;YAN Yang(School of Computer Science and Technology,Changchun Normal University,Changchun 130032,China)

机构地区：[1]长春师范大学计算机科学与技术学院,吉林长春130032

出　　处：《长春师范大学学报》2022年第12期62-67,73,共7页Journal of Changchun Normal University

基　　金：吉林省教育厅科学技术研究项目“基于深度学习的眼底图像视盘自动定位方法研究”(JJKH20210887KJ);吉林省自然科学基金资助项目“基于人工智能的眼底病相关目标精准检测关键技术研究及应用”(YDZJ202101ZYTS147)。

摘　　要：针对视网膜血管结构复杂、图像对比度低与细节区域分割不精准问题,提出一种基于改进U-Net分割算法。针对卷积操作时卷积核的感受野范围较小而不能充分提取血管特征的问题,将原始卷积层替换成可变形卷积模块,该模块组合了不同尺度、不同复杂度的分支来丰富特征空间的多样性,增大卷积核的感受野范围,进而提升血管特征提取的效果;针对采样操作时产生的梯度消失问题,在网络上采样的过程引入循环残差卷积模块,有助于训练深层网络架构,解决梯度消失问题,避免冗余特征影响,更好地表示图像特征。将本文方法在DRIVE数据集上进行数据对比实验,实验结果的准确性为95.59%,特异性为97.92%,敏感性为79.63%,与当前主流的视网膜血管分割方法相比,改进的模型性能具有一定优势。To address the problems of complex retinal vascular structure,low image contrast and imprecise segmentation of detailed regions,an improved U-Net segmentation algorithm is proposed.To address the problem that the perceptual field of the convolution kernel is small during the convolution operation,the original convolution layer is replaced by a deformable convolution module,which combines branches of different scales and complexities to enrich the diversity of the feature space and increase the perceptual field of the convolution kernel,thus improving the effect of vascular feature extraction.To address the gradient disappearance problem generated during the sampling operation,a cyclic residual convolution module is introduced in the process of sampling on the network,which helps to train the deep network architecture,solve the gradient disappearance problem,avoid the influence of redundant features,and better represent the image features.The data comparison experiments of this paper’s method on the DRIVE dataset showed an accuracy of 95.59%,a specificity of 97.92%and a sensitivity of 79.63%,and the performance of the improved model has certain advantages compared with the current mainstream retinal vessel segmentation methods.

关 键 词：血管分割 U-Net模型 可变形卷积 循环残差卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]