【作者】 陈宁华；

【导师】 张三元；

【作者基本信息】 浙江大学 ， 工业设计工程， 2017， 硕士

【摘要】 随着人们生活质量的不断提高,糖尿病患者也呈现逐年增多的趋势。糖尿病性视网膜病变作为糖尿病性微血管病变的严重并发症之一,是引起失明以及视力障碍的主要原因。这种病理信息往往可以从视网膜图像中分析得出,然而对于不同的病变阶段,视网膜图像会呈现出不同的病理构造;此外,视网膜图像中的微血管系统也因人而异。这些因素都给本文的研究带来巨大的挑战。本文以视网膜图像作为研究对象,着重对视网膜图像的解剖结构以及病变区域等内容进行分析,主要的研究内容分为以下三个方面:1、视网膜图像预处理。此阶段的主要任务包含视网膜图像背景均衡化,图像噪声去除和感兴趣区域提取三个方面。预处理工作的好坏直接关系到解剖结构检测和病变分析的准确性。2、视网膜图像解剖结构检测。包括血管分割、视神经盘检测以及黄斑区域定位。在视网膜血管分割上,提出一种基于韦伯变换和多尺度空间分析的方法,并通过边界强度和脊强度这两个重要特征对血管像素进行增强;视神经盘作为青光眼神经损害的诊断依据,本文利用其本质的特征并结合Sobel算子、Canny算子以及霍夫变换得到最终的检测结果;对于黄斑区域的工作,文中主要应用模板匹配的方法对其实现精确定位。3、视网膜图像病变区域检测。主要检测的对象为微动脉瘤、出血以及渗出。在微动脉瘤的检测上,运用多尺度高斯核函数进行相似性匹配,接着通过支持向量机对获得的检测点进行分类,该方法的准确率较其他算法相比有着明显的提升。对于视网膜图像中的出血检测,将基于多尺度线性结构的数学形态学方法检测得到的候选区域,通过SLIC算法处理成超像素块并计算其分类特征,利用灰度值投票算法进行分类。最后,针对渗出的检测,先是通过滑动窗口以及形态学重建的方法获得渗出候选区域,然后根据超像素区域计算其对应的特征,并设计人工神经网络分类器进行分类。更多还原

【Abstract】 With the continuous improvement of people’s quality of life,the incidence of diabetes show increasing trend year by year.Diabetes retinopathy(DR)is one of the major complications associate with diabetic microvascular disease and a main reason causes blindness and visual impairment.The pathological information of DR can be analyzed from the retinal image.Due to the diverse pathologic structures of different disease stages and the uniqueness of human retinal micro-vascular system,we will confront an extremely tough challenge when taking the research.In this paper,retinal images are used as objects of study and the major work focuses on the analysis of anatomical structure and lesion regions.The main research contents are listed as follows:1.Retinal image preprocessing.This task contains retinal image background homogenization,noise pixel attenuation and region of interest extraction.And the result of preprocessing has an effect on the accuracy of anatomic structure detection and lesion analysis.2.Retinal anatomical structure detection.It includes vascular segmentation,optic disc detection and macular localization.We propose a method based on Weber transformation and multi-scale spatial analysis to segment vessels.Then enhance the vessel pixels through two key features,edge strength and ridge strength.Optic disc(OD)acts as the diagnostic basis of glaucomatous optic nerve damage,we can use its essential characteristics combined with Sobel operator,Canny operator and Hough transformation to get the final detected OD.And as for macular localization,an approach of template matching is used for getting the precise location.3.Retinal lesion regions detection.The main detecting target consists of micro-aneurysms,hemorrhages and exudates.In the detection of micro-aneurysms,a multi-scale Gaussian kernel function is used for similarity matching,and then employing Support Vector Machine(SVM)to classify the collected candidates.Compared with other methods,the proposed method has a higher accuracy.As for hemorrhages,the candidate regions are obtained by the mathematical morphology with multi-scale linear structure.And then process the regions into super-pixel blocks by using SLIC and calculate its classification features,which are used for classifying hemorrhages by means of grayscale voting algorithm.The last part lies in exudates detection,we first use dynamic window and morphological reconstruction to get the candidate regions.Then the features are calculated corresponding to the super-pixel blocks,and design an Artificial Neural Networks(ANN)classifier.更多还原