摘要
视网膜血管病变程度可作为与血管相关疾病的诊断依据,因此对视网膜血管的特征提取和分析在疾病的诊断、治疗上具有重要的临床价值.然而视网膜血管图像结构复杂,血管与背景存在灰度值交叉以及病灶和噪声的影响,致使诸多算法在相邻血管处、血管交叉处和微血管处存在分割不足的问题,且易受病灶和噪声的干扰.针对现有血管分割方法的缺陷,同时考虑视网膜图像局部血管结构与背景的可分性以及模型的抗噪性,文中提出了一种融合区域能量拟合信息和形状先验的水平集血管分割方法.在图像预处理过程中,利用GAC水平集模型控制轮廓曲线朝目标真实边界演化来获得视网膜掩模,接着通过形态学算子去除血管中心亮线以及利用高斯卷积估计视网膜背景图像,并与原图像和掩模分别进行减法和点乘运算增强视网膜血管图像;然后分析Hessian矩阵的各向异性特性,它的特征值在血管、背景和病灶上具有不同的几何性质,为最大化这种几何性质在不同结构上的差异,文中利用Hessian矩阵特征值重新构建血管响应函数,从而获得视网膜血管初步图像,作为后面的先验信息、初始化信息等,缓解水平集对初始化敏感和易误分割的问题.在水平集血管模型构建中,考虑到视网膜图像局部血管的可分性、模型的抗噪性等因素,使用RSF模型的局部区域能量拟合性质划分血管与背景,以及利用先验信息和水平集函数的几何性质构建水平集模型的形状约束项,当轮廓曲线远离视网膜血管图像的先验信息位置时就会受到惩罚,距离越远惩罚越大.完整的水平集血管分割模型包含了局部区域能量拟合项、形状约束项、面积约束项等,在能量最小化求解过程中,能够通过局部区域能量拟合性质克服现有血管分割方法在相邻血管处、血管交叉处和微血管处分割不足的问题,以及通过形状约束性质提高模型的抗噪性;最后,利用连通域面积和宽、高信息构建几何算子,进一步消除连通域面积较小的伪影和病灶,获得最终的视网膜血管分割图像.通过在以下三个数据库上的实验仿真,对HRF数据库的HRF_healthy、HRF_diabetic和HRF_glaucoma眼底图像数据,敏感度分别达到79.4547%、81.0653%、81.1773%,准确率分别达到96.1820%、94.2147%、95.6413%;对STARE、DRIVE数据库上的眼底图像数据敏感度分别达到79.0860%、75.3535%,准确率分别达到95.0340%、95.3565%.
Thanks to the severity of retinal vessel lesions is an important reference to the diagnosis of vessel diseases;the segmentation and analysis of the retinal vessel have a great clinic value in diagnosing and treating these diseases.However,due to complex structure of the retinal blood vessel,gray level crossing and the influences of the focus and noise,many algorithms have disadvantage of inadequate segmentation on the position of adjacent vessels,crossing vessels and micrangium as well as being affected easily by the focus and noise.Aiming to the shortcomings of existed methods and considering that partial vessel structure can be separated from background and the resistance of model to noise,we propose level set method to segment blood vessel fusing regional energy fitting information and shape priori.First,in the stage of preprocessing,it uses GAC level set model to control contour curve evolving to real border,which leads to retinal mask,eliminates central bright line of vessel and evaluates background image of retina,conducts subtraction and multiplication with initial image and mask to reinforce retinal vessel image.Than by the characteristic analysis of anisotropy of Hessian,its character values have different geometric property on vessels,background and the focus,to maximize the difference on different structures,we rebuild vessel response function by the eigvalues of Hessian matrix to acquire primary image of retina as prior and initialization information thereafter,which alleviate the sensitivity of level set model to initialization and missegmentation.On constructing the model of level set,taking into consideration separation of local retinal image,noise-resistance of model and so forth,we make a distinction between vessels and background via the property of regional energy fitting of RSF model and build shape constraints of level set using prior information as well as geometric property of level set function,the curve will receive punishment when it is away from the position of prior information,the further,the bigger the punishment is.The integral vessel segmentation model of level set is composed of local energy fitting,shape constraint,area constraint;on the process of minimization,it overcome the shortcomings by energy fitting of insufficient segmentation at adjacent vessels,microvascular and vascular intersection that exist in many established method,and improve the noise-resistance of model via shape constraint.At last,the final retinal segmentation image is achieved by area of connected domain and geometric operator constructed by width and height information which can further eliminate the fake shadow of connected domain and the focus.By the experiments on following three databases,the sensitivity is 79.4547%,81.0653%,81.1773%,accuracy is 96.1820%,94.2147%,95.6413%corresponding to dataset of fundus image of HRF_healthy,HRF_diabetic and HRF_glaucoma in database of HRF;on STARE,DRIVE,the sensitivity is 79.0860%,75.3535%,accuracy is 95.0340%,95.3565%respectively.The results above show the effectiveness of vessel segmentation method proposed in this paper.
作者
梁礼明
黄朝林
石霏
吴健
江弘九
陈新建
LIANG Li-Ming;HUANG Zhao-Lin;SHI Fei;WU Jian;JIANG Hong-Jiu;CHEN Xin-Jian(School of Electronic and Engineering and Autom ation,Jiangxi University of Science and Technology,Ganzhou,Jiangxi 341000;School of Electronics and Inform ation Engineering,Soochow University,Suzhou,Jiangsu 215006)
出处
《计算机学报》
EI
CSCD
北大核心
2018年第7期1678-1692,共15页
Chinese Journal of Computers
基金
国家"九七三"重点基础研究发展规划项目基金(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(20132BAB203020)
江西省教育厅科学技术研究重点项目(GJJ170491)资助~~