一种有效的视网膜图像血管检测算法.doc

AN EFFICIENT BLOOD VESSEL DETECTION ALGORITHM FOR RETINAL IMAGES一种有效的视网膜图像血管检测算法USING LOCAL ENTROPY THRESHOLDING利用局部熵阈值Thitiporn Chanwimaluang and Guoliang FanSchool of Electrical and Computer Engineering电气与计算机工程学院Oklahoma State University, Stillwater, OK 74078美国俄克拉荷马州立大学，静水，好74078Email: thitipo,电子邮件： thitipo，glfan ABSTRACT 摘要This paper presents an efficient method for automatic detection and extraction of blood vessels in retinal images. Specifically, we also delineate vascular intersectionslcrossovers. The proposed al-gorithm is composed of four steps: matched filtering, local entropy thresholding, length filtering, and vascular intersection detection. The purpose of matched filtering is to enhance the blood vessels. Entropy-based thresholding can well keep the spatial structure of vascular tree segments. Length filtering is used to remove mis-classified pixels. The algorithm has been tested on twenty ocular fundus images, and experimental results are compared with those obtained from a state-of-the-art method and hand-labeled ground truth segmentations. 本文提出了一种有效的方法，用于自动检测和提取的血管在视网膜图像。具体来说，我们也划定血管intersectionslcrossovers。建议的铝去算法包括四个步骤：匹配滤波、局部熵的阈值，长度过滤，和血管的交叉点检测。匹配滤波的目的是增强血管。基于熵的阈值化能较好地保持血管树段的空间结构。长度滤波是用来删除错误分类的像素。二十眼基金的算法进行了测试我们的图像，实验结果与一个国家的最先进的方法和手工标记的地面真理分割得到的比较。1. INTRODUCTION1。简介The automatic detection of blood vessels in the retinal images can视网膜图像中血管的自动检测help physicians for the purposes of diagnosing ocular diseases, pa-tient screening, and clinical study, etc. Information about blood帮助医生诊断眼部疾病的目的，患者的筛查和临床研究，对血液等信息vessels in retinal images can be used in grading disease sever-ity or as part of the process of automated diagnosis of diseases. Blood vessel appearance can provide information on pathological changes caused by some diseases including diabetes, hypertension, and arteriosclerosis. The most effective treatment for many eye-related diseases is the early detection through regular screenings. Furthermore, a segmentation of the vascular tree seems to be the most appropriate representation for the image registration applica-tions due to three following reasons: I ) it maps the whole retina; 2) it does not move except in few diseases; 3) i t contains enough information for the localization of some anchor points.血管在视网膜图像可用于疾病严重度的分级或疾病的自动诊断过程的一部分。血管外观可以提供病理改变的信息由某些疾病引起的糖尿病，高血压和动脉硬化。最有效的治疗许多眼部相关疾病是通过定期检查的早期检测。而且再次，一个分割的血管树似乎是为图像配准应用中，由于以下三个原因最合适的表示：我）它映射整个视网膜；2）它只的不动，除了在一些疾病；3）我没有足够的信息，一些锚点定位。There are many previous works on extracting blood vessels in retinal images. In edge detection-based method I , since lo-cal gradient maxima occur at the boundary of the vessels, the sig-nificant edges along these boundaries are extracted. The group-ing process searches a partner for each edge which satisfies cer-tain criteria like opposite gradient direction and spatial proximity. In tracking-based method Z, each vessel segment is defined by three attributes. direction, width, and center point. The density distribution of cross section of a blood vessel can be estimated us-ing Gaussian shaped function. Individual segments are identified using a search procedure which keeps track of the center of the vessel and makes some decisions about the future path of the ves-sel based on certain vessel properties. This method requires that beginning and ending search points are manually selected using cursor. An efficient piecewise threshold probing technique was proposed in 3 where the matched-filter-response (MFR) image is有许多以前的作品在视网膜图像中提取血管。在边缘检测方法我，因为局部梯度最大值出现在血管的边界，产生的重要了GES沿着这些边界提取。集团的合作伙伴搜索过程的每个边缘，满足一定的标准像相反的梯度方向和空间距离。在tracking-b本文方法Z，各血管段是由三个属性定义。方向，宽度和中心点。血管的横截面的密度分布，可以估计我们荷兰模仿功能。个体段使用一个搜索过程跟踪容器中心的确定和对未来的一些决定VES路径选择基于一定的容器性能。此方法要求使用游标手动选择开始和结束搜索点。在 3 中提出了一种有效的分段阈值探测技术子滤波器响应（MFR）图像used for mapping the vascular tree. A set of criteria is tested to de-termine the threshold of the probe region, and ultimately to decide i f the area being probed is a blood vessel. Since the MFR image is probed in a spatially adaptive way, different thresholds can be applied throughout the image for mapping blood vessels.用于绘制血管树。一套标准的测试确定探测区域的阈值，并最终决定我F区正在探索的是一种血管。由于MFR IMA通用电气的空间自适应方式，不同的阈值可以应用于整个图像映射血管。In this paper, we propose a new algorithm to efficiently locate and extract blood vessels in ocular fundus images. The proposed algorilhm is composed of four steps, matched filtering, entropy-based thresholding, length filtering, and vascular intersection de-tection. Compare with the method in Z. our proposed algorithm does not involve human intervention. Since our algorithm can au-tomatically estimate one optimal threshold value, i t requires less computational complexity compared with the method in (31.在本文中，我们提出了一种新的算法来有效地定位和提取眼底图像的血管。该算法分为四个步骤，匹配滤波，基于熵的日resholding，长度过滤，和血管交叉点检测。与的方法比较。我们所提出的算法不涉及人为干预。由于我们的算法可以自动E估计一个最优阈值，我不需要较少的计算复杂性的方法相比（31。2. PROPOSED ALGORITHM 2。算法The proposed algorithm is composed of four steps. Since blood vessels usually have lower reflectance compared with the back-ground, we apply the matched filter to enhance blood vessels with the generation of a MFR image. Secondly, an entropy-based thresh-olding scheme can be used to distinguish between vessel segments and the background in the MFR image. A length filtering tech-nique is used to remove misclassified pixels. Vascular intersection detection is performed by a window-based probing process. 所提出的算法是由四个步骤。由于血管通常有较低的反射率相比，我们采用匹配的过滤器，以增强血管的属一个MFR图像化。其次，基于熵的阈值方案可以用来区分血管段和MFR图像背景之间。一个长度过滤技术是用来重新移动的错误分类的像素。血管交叉检测的窗口为基础的探测过程。2.1. Matched Filter 2.1。匹配滤波器In 4. the gray-level profile of the cross section of a blood vessel can be approximated by a Gaussian shaped curve. The concept of matched filter detection is used to detect piecewise linear segments of blood vessels in retinal images. Blood vessels usually have poor local contrast. The two-dimensional matched filter kernel is de-signed to convolve with the original image in order to enhance the blood vessels. A prototype matched filter kernel is expressed as 在 4 。血管的横截面的灰度曲线可以近似为高斯形曲线。匹配滤波器检测的概念是用来检测分段线性段视网膜图像中的血管。血管通常有较差的局部对比度。二维匹配滤波器内核设计了卷积以原始图像增强血管。一个原型匹配滤波器内核表示为f ( Z , Y ) = -e.p(&), for IYI 5 L P , (1) F（z，y）= -（&），为的iyi 5 L P，（1）where L is the length of the segment for which the vessel is as-sumed to have a fixed orientation. Here the direction of the vessel is assumed to be aligned along the y-axis. Because a vessel may be oriented at any angles, the kernel needs to be rotated for all possible angles. A set of twelve 16x15 pixel kernels is applied by convolving to a fundus image and at each pixel only the max-imum of their responses is retained. For example, given a retinal image in Fig. 2(a) which has low contrast between blood vessels and background, its MFR version is shown in Fig. 2(b), where we can see blood vessels are significantly enhanced. 其中L是长度的段，该船是假设有一个固定的方向。这里的船的方向是沿Y轴对齐。因为一艘船可能是一种在任何角度，内核需要对所有可能的角度旋转。一套十二16x15像素内核采用卷积到眼底图像，每个像素只最大的红外响应保留。例如，在图2中给出的视网膜图像（一）具有血管和背景之间的对比度低，其MFR版本显示在图2（b），在那里我们可以看到血管SelS显著增强。2.2Local Entropy Thresholding Secondly, the MFR image is processed by a proper thresholding scheme in order to extract the vessel segments from the back- ground. An efficient entropy-based thresholding algorithm. which takes into account the spatial distribution of gray levels, is used be- cause an image pixel intensities are not independent of each other. Specifically, we implement a local entropy thresholding technique, described in 5 which can well preserve the spatial structures in the binarized/thresholded image. 7uo images with identical his- tograms but different spatial distribution will result in different en- tropy (also different threshold values). The co-occurrence matrix of the image F is an P x Q dimen- sional matrix T = tijpxQ that gives an idea about the transi- tion of intensities between adjacent pixels, indicating spatial struc- tural information of an image. Depending upon the ways in which the gray level i follows gray level J , different definitions of co- occurrence matrix are possible. Here, we made the co-occurrence matrix asymmetric by considering the horizontally right and verti- cally lower transitions. Thus, t; is defined as follows: 2.2local熵阈值其次，MFR图像是由一个适当的门限设置为从后提取血管段地面处理。一个高效的基于熵的阈值算法。这考虑到的灰度级的空间分布，被使用是-导致一个图像的像素强度是不独立的彼此。具体来说，我们实现了一个局部熵Y的阈值方法，描述 5 可以很好的保持空间结构/阈值的图像二值化。同他的图谱的不同空间分布7uo图像会导致不同的EN -异性（也有不同的阈值）。灰度共生矩阵的图像F是P x Q维维矩阵T = 兰 PXQ给一个想法的过渡、Of强度相邻像素之间的空间结构，显示的图像结构信息。根据其灰度级灰度，我遵循的方式，共同定义不同OC矩阵是可能发生。在这里，我们的共生矩阵不对称考虑水平和垂直上下的转换。因此，定义如下：(2)The probability of co-occurrence pij of gray levels i and j can 的共现的概率Pij灰度i和j可以therefore be written as 因此被写(3)If s. 0 5 s 5 L - 1, is a threshold. Then s can partition 如果美国5 0的5升- 1，是一个门槛。然后可分区the cohccurrence matrix into 4 quadrants, namely A, B. C, and D 的cohccurrence矩阵分为4个象限，即A、B、C、DFigure 1: Quadrants of co-occurrence matrix 5 图1：共生矩阵 5 象限-1 Let us define the following quantities: 让我们来定义以下数量：(4)Normalizing the probabilities within each individual quadrant, 在每个单独的象限内规范概率，such that the sum of the probabilities of each quadrant equals one, 这样，每个象限的概率之和等于一，we get the following cell probabilities for different quadrants: 我们得到以下细胞概率不同象限：(5)The second-order entropy of the object can be defined as 对象的二阶熵可以被定义为Similarly, the second-order entropy of the background can be written as类似地，二阶熵的背景可以写作为Hence, the total second-order local entropy of the object and the 因此，总的二阶局部熵的对象和background can be written as 背景可以写成The gray level corresponding to the maximum of H$(s) gives 对应于最高的“$”（秒）的灰度级the optimal threshold for object-background classification. For the 目标背景分类的最佳阈值。为MFR image shown in Fig. 2(b), the entropy-based thresholding MFR图像如图2（b），基于熵的阈值result is shown in Fig. 2(c) where we can see blood vessels are 结果如图2所示，我们可以看到血管的出现clearly segmented from the background 从背景中清晰地分割2.3. Length Filtering 2.3。长度滤波As seen in Fig. 2(c), there are still some misclassified pixels in the 如图2（c），仍然有一些错误分类的像素在image. Here we want to produce a clean and complete vascular 图像。在这里，我们要生产一个干净和完整的血管tree structure by removing misclassified pixels. Length filtering is 通过去除错误分类的像素结构树。长度滤波used to remove isolated pixels by using the concept of connected 用连通的概念来去除孤立像素pixels labeling. Connected regions correspond to individual objects. 像素标记。连通区域对应于单个对象。We first need to identify separate connected regions. The 我们首先需要确定单独的连接区域。这个length filtering tries to isolate the individual objects by using the 长度过滤试图通过使用该方法来隔离单个对象eight-connected neighborhood and labepropagation. Once the algorithm 八连通邻域和标签传播。算法一次is completed, only the resulting classes exceed a certain 完成后，只产生类超过某个number of pixels, e.g., 250, are labeled as blood vessels. Fig. 2(d) 像素数，例如，250，被标记为血管。图2（4）shows the results after length filtering. 显示长度滤波后的结果。2.4. Detection of Vascular IntersectionslCrossovers 2.4。血管intersectionslcrossovers检测Vascular intersections and crossovers are the most appropriate representation 血管交叉点和交叉是最合适的表示in registration process because they exist in every retinal 在注册过程中，因为它们存在于每一个视网膜images, and do not move except in some diseases. If a vascular 图像，不要动，除了在某些疾病。如果血管tree is one-pixel wide, the branching points can be detected and 树是一个像素宽，分支点可以被检测到characterized efficiently from the vascular tree. Morphological 有效地从血管树的特点。形态thinning is applied to the vascular tree in order to get one-pixelwide 细化应用于血管树为了得到一pixelwidevascular tree as shown in Fig. 2(e). In order to save computational 血管树如图2所示（电子）。为了节省计算time, a 3 x 3 neighborhood window is used to probe and find 时间，33邻域窗口被用来探测和查找the branching points. If the number of vascular tree in the window 分支点。如果窗户里血管树的数目is great than 3, it is a branch point. Then a 11 x 11 neighborhood is 是伟大的3，它是一个分支点。然后一个1111邻域是applied through a detected branching points in order to eliminate 应用于检测分支点以消除the small intersections 6. We consider only the boundary pixels 小十字路口 6 。我们只考虑边界像素of a 11 x 11 square. If the number of vascular tree on the boundary 一个1111平方米。如果边界上的血管树的数目is greater than 2, we mark it as an intersection/crossover. Fig. 2(0 是大于2，我们作为一个交叉/交叉标记。图0（2presents the vascular tree with the intersections and crossovers. 介绍了血管树的交叉点和交叉。 Figure 2: (a) An original retinal image. (b) Matched filtering result. (c) Local entropy thresholding result. (d) Vascular tree. (e) One-pixel 图2：（一）原始的视网膜图像。（乙）匹配滤波结果。局部熵阈值结果。（维）血管树。（一）一像素wide vascular tree. (0 One-pixel wide vascular tree with intersections and crossovers overlaying on gray-scaled image. 宽维血管树。（0个像素宽的血管树的交叉点和交叉的灰度图像。3. SIMULATION RESULTS 3。仿真结果On Windows XP. Pentium 4, CPU 1.7 GHz. usinr MATLAB verrithm, 在Windows XP。奔腾4，中央处理器1.7。usinr MATLAB verrithm，which takes into account the spatial distribution ofgray levels, 考虑到空间分布的灰度级，performs efficiently in distinguishing between enhanced vessel 有效地进行增强的血管之间的区别seements and the backeround since it can meserve the structurede- sion 6.1, the computational time of the whole process of our algorithm seements和低音调子组成了背景音乐因为它可以莫瑟夫的structurede - 锡永6.1，我们对整个算法过程的计算时间takes approximately 3 minutes for each retinal image. We 为每一个视网膜图像大约需要3分钟。我们use the same set of twenty 605 x 700 pixel retinal images (24bpp). 用二十个605700像素的视网膜图像相同的设置（24bpp）。as used in 3. In order to evaluate the performance of our algorithm, 如用在 3 。为了评估我们的算法的性能，we compare our simulation results with the state-of-the-art 我们比较我们的模拟结果与国家的最先进的results obtained from 3 and hand-labeled groundtruth segmentations 从 3 和手工标记groundtruth分割得到的结果as shown in Fig. 3. 如图3所示。Specifically, we classify retinal images into three categories. 具体而言，我们将视网膜图像分为三类。normal retinal images, abnormal retinal images with some lesions, 正常视网膜图像，部分病变视网膜病变，and retinal images with obscure blood vessel appearance. From 与不明血管外观的视网膜图像。从column by column. the first column of Fig. 3 presents the results 柱柱。图3的第一列给出的结果from normal retinal images. Results from abnormal retinal images 从正常视网膜图像。异常视网膜图像的结果with some lesions are shown on the second column. The last column 在第二柱上显示病变。最后一列presents results from retinal images with obscure blood vessel 从视网膜图像的模糊血管呈现的结果appearance. From row by row, the first row shows original retinal 外观。连续行，第一行显示原视网膜1 1“tails of an image. However, the presence of lesions in the abnormal 图像的尾巴。然而，在异常病变的存在retinal image is the major obstacle in extracting blood vessels since 视网膜图像是提取血管的主要障碍they are also mis-enhanced and mis-detected as blood vessels. Our 他们也被错误的增强和错误检测为血管。我们的algorithm is sensitive to lesions due to the fact that their boundaries 算法是敏感的病变，由于这样的事实，它们的边界partially match the shape of matched filter kernels, while the algorithm 部分匹配的匹配滤波器的内核的形状，而算法in 3 is more robust to lesions. We expect to improve the 在 3 是更强大的病变。我们希望改善robustness of our algorithm by involving color information and 我们的算法的鲁棒性，涉及颜色信息和additional anatomical constraints for blood vessel detection and 血管检测的附加解剖限制extraction. 提取。4. CONCLUSIONS 4。结论In this paper, we have introduced an efficient algorithm for fully 在本文中，我们已经介绍了一种有效的算法充分automated blood vessel detection in ocular fundus images using 眼底图像中血管的自动检测images. The second row-shows the hand-labeled groundtruth segmentations 图像。第二行显示手标记groundtruth分割which were manually labeled by Hoover 3. The third 这是手工标记的胡佛 3 。第三row presents simulation results from 31. The last row presents 行给出了仿真结果。最后一排礼物our simulation results. Although, algorithm in 3 performs very 我们的模拟结果。虽然，算法在 3 执行非常well, a significant improvement can be achieved by our algorithm 一个显着的改善，可以实现我们的算法for normal and obscure retinal images. Our method performs very 对于正常和不起眼的视网膜图像。我们的方法执行非常well in extracting blood vessels. Even the smaller blood vessels 提取血管的好。甚至更小的血管can he extracted. Matched filtering enhances the contrast of blood 他可以提取。匹配滤波提高了血液的对比度vessels against the background. Local entropy thresholding algothe 背景血管。局部熵阈值algothelocal entropy thresholding scheme. The proposed method re- 局部熵阈值法。所提出的方法重新-tains thecomputational simplicity, and at the same time, can achieve 从计算简单，同时，可以实现accurate segmentation results in the case of normal retinal images 正常视网膜图像的精确分割结果and images with obscure blood vessel appearance. In the case of 与不明血管外观的图像。在这种情况下abnormal retinal images with lessons, some lesions are also misdetected 异常的视网膜图像的教训，一些病变也误检测in addition to blood vessels. In the future work, we want 除血管外。在未来的工作中，我们要to improve the robustness of our algorithm by involving in the preprocessing 在预处理中提高算法的鲁棒性scheme and additional anatomical constraints to separate 计划和额外的解剖约束分开the lesions in the final vascular tree. 最后血管树的病变。Figure 3: First row: Example images; Second row: Hand-labeled