基于多特征融合的有监督视网膜血管提取  

Supervised Blood Vessel Extraction in Retinal Images Based on Multiple Feature Fusion

作　　者：梁礼明 刘博文 杨海龙 石霏[3] 陈新建[3] LIANG Li-Ming;LIU Bo-Wen;YANG Hai-Long;SHI Fei;CHEN Xin-Jian(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou,Jiangxi 341000;State Grid Guangyuan Power Supply Company,Guangyuan,Sichuan 628000;School of Electronics and Information Engineering,Soochow University,Suzhou,Jiangsu 215006)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000 [2]国家电网广元供电公司,四川广元628000 [3]苏州大学电子信息学院,江苏苏州215006

出　　处：《计算机学报》2018年第11期2566-2580,共15页Chinese Journal of Computers

基　　金：国家"九七三"重点基础研究发展计划项目(2014CB748600);国家自然科学基金(81371629;61401293;61401294;81401451;81401472;51365017);江苏省自然科学基金(BK20140052);江西省自然科学基金资助项目(20132BAB203020);江西省教育厅科学技术研究重点项目(GJJ170491)资助~~

摘　　要：视网膜血管提取在眼科疾病的诊断和治疗中具有重要的临床价值,但由于其拓扑结构复杂与病灶噪声等原因,现有的提取方法精度低、鲁棒性差.为此,该文提出了一种基于多特征融合的有监督学习的视网膜血管提取方法.首先提取视网膜血管的线性特征、纹理特征、矩特征、方差特征和灰度特征等作为样本特征;然后通过随机森林模型训练得到视网膜图像血管分类器,由随机森林算法初步提取血管;最后利用视网膜血管灰度分布信息和连通域信息进行血管图像后处理,进一步去除初步提取结果中的伪影和病灶等非血管成分,获得最终的视网膜血管分割图像.通过在DRIVE和STARE眼底图像数据库上的实验仿真可知,该文算法的敏感度分别为0.8354、0.8452,准确率分别为94.83%、95.34%,总体指标优于已有的视网膜血管图像分割方法.Retinal blood vessel segmentation plays an important role in ophthalmic diagnosis and treatment.Many research achieved segmentation with single feature extraction in decades,including supervised and unsupervised methods.Despite the results of these works were acceptable and significant,retinal vessel would be rupture or too thin in width on final binary images.To solve the above problem,this paper presents a supervised method for blood vessel extraction based on multiple feature fusion in retinal images.The proposed method consists of four steps.First,green channel images are separated from raw retinal images for their clear blood vessel information.Optic disc and macular center are removed by bilateral filtering.The intensity of processed images is stretched to 0-255.Second,according to changes in vessel morphology such as shape,tortuosity,branching pattern,width and size,a 34-D feature vector is extracted from linear feature,texture feature,moment feature,variance feature and intensity feature.Most retinal vascular morphological features are taken into consideration by means of the multiple feature fusion vectors.These features describe retinal vessel from edge information,intensity difference,blood vessel size,connectivity and statistics.Third,feature vectors in every pixel are treated as training data and manual segmentation standard is treated as training label.A random forest classifier,known for its simplicity,anti-overfitting and information fusion capability,is trained with 34-D feature vector and its label.Taken different feature into consideration,100 decision trees are established from random forest model.During each decision tree,the vessel pixel is set as positive point and the non-vessel pixel is set as negative point.Preliminary binary segmentation results are voted by multi-decision tree classifier.Finally,segmentation results are obtained from vessel image post-processing.Small or non-vessel pixels,like artifact and lesion noise,are removed via vessel image post-processing filter operators which

关 键 词：视网膜血管提取 特征提取 有监督学习 随机森林 血管图像后处理 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]