视网膜血管图像分割算法的研究

发布时间：2018-01-29 06:14

本文关键词： 视网膜血管分割局部增强全局增强 Zernike矩多尺度多模板　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：人眼视网膜血管形态结构.与多种疾病有着非常密切的联系,能够为眼科疾病以及全身性系统疾病提供前期诊断和预防的依据。传统的人工筛查,不但工作量大,花费大量时间,而且准确率较低,并有可能造成误诊。而随着数字图像处理技术在医学图像领域的发展以及计算机的广泛应用,视网膜血管结构的自动检测已成为一种趋势。目前研究者们提出了很多视网膜血管的自动分割算法,取得一定成效,但是对血管分割的准确率及其视网膜自身组织结构的影响有待进一步提高。基于此问题,本文提出两种新的视网膜血管分割算法,并在STARE和DRIVE两个图像库中进行大量实验,最后将实验结果和已知的分割算法进行比较,论文主要工作如下:1)针对与背景像素对比度较低的细小血管,其分割准确率低,造成细小血管的丢失以及血管断裂等问题,本文提出了一种基于多尺度的局部与全局增强相结合的视网膜血管分割算法。首先将多尺度线性检测器进行再划分,分为小尺度和大尺度两部分;其次对小尺度下的图像进行局部增强和对大尺度下的图像进行全局增强处理;再次将小尺度下的响应函数以及大尺度下的响应函数与其分别对应的增强图像进行线性融合;最终对获得的融合图像进行后续处理,去掉非连通区域和孤立点,得到最终的视网膜血管分割图像。理论分析和实验结果表明,本文算法分割出更多的细小血管,在STARE和DRIVE两个图像库中的准确率分别达到0.9662和0.9645,高于传统分割算法。2)针对因视盘与视网膜血管结构相似而造成视盘被误分割为视网膜血管的问题,本文提出了一种基于Zernike矩不同模板系数的视网膜血管分割算法。首先计算Zernike矩3×3、5×5、7×7、9×9四个不同的矩模板系数M20,并研究不同模板系数对不同宽度血管的影响;其次对上述四个模板卷积结果进行多尺度线检测,得到各模板下的响应函数;再次将各自的响应函数与其分别对应的矩图像进行线性融合;最后对获得的融合图像进行后续处理,去掉非连通区域和孤立点,得到最终的视网膜血管图像。理论分析和实验结果表明,本文算法在STARE和DRIVE两个图像库中的准确率分别达到0.9659和0.9643,且能够对视盘进行有效处理。
[Abstract]:The morphological structure of retinal vessels in human eyes is closely related to many diseases and can provide the basis for early diagnosis and prevention of ophthalmic diseases and systemic diseases. Traditional artificial screening. With the development of digital image processing technology in medical image field and the wide application of computer, not only the workload is large, but also the accuracy rate is low, and it may cause misdiagnosis. Automatic detection of retinal vascular structure has become a trend. At present, researchers have proposed a lot of automatic retinal blood vessels segmentation algorithm, and achieved certain results. However, the accuracy of blood vessel segmentation and the influence of retinal tissue structure need to be further improved. Based on this problem, two new retinal vascular segmentation algorithms are proposed in this paper. A large number of experiments are carried out in the STARE and DRIVE image databases. Finally, the experimental results are compared with the known segmentation algorithms. The main work of this paper is as follows: (1) aiming at small blood vessels with low contrast with background pixels, the segmentation accuracy is low, resulting in the loss of small vessels and vascular breakage. In this paper, a multi-scale local and global enhancement based retinal vascular segmentation algorithm is proposed. Firstly, the multi-scale linear detector is subdivided into two parts: small scale and large scale. Secondly, local enhancement of small scale image and global enhancement of large scale image are carried out. Finally, the response function of small scale and the response function of large scale are fused linearly with the corresponding enhancement image. Finally, the fusion image is processed to remove the disconnected region and the isolated point, and the final retinal vascular segmentation image is obtained. The theoretical analysis and experimental results show that. In this paper, more small blood vessels are segmented, and the accuracy in STARE and DRIVE image database is 0.9662 and 0.9645, respectively. Higher than the traditional segmentation algorithm. 2) aiming at the problem that the optic disc is divided into retinal vessels by mistake because of the similarity between the optic disc and the retinal vascular structure. In this paper, a retinal blood vessel segmentation algorithm based on different template coefficients of Zernike moments is proposed. Firstly, the Zernike moment 3 脳 3 脳 5 脳 5 脳 5 脳 7 is calculated. Nine 脳 9 four different moment template coefficients M20, and the effects of different template coefficients on different width of blood vessels were studied. Secondly, the multi-scale line detection is carried out on the convolution results of the above four templates, and the response function under each template is obtained. Thirdly, the corresponding moment images are fused linearly with their respective response functions. Finally, the fusion image is processed to remove the disconnected region and isolated point, and the final retinal vascular image is obtained. The theoretical analysis and experimental results show that. The accuracy of this algorithm in STARE and DRIVE image database is 0.9659 and 0.9643 respectively.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R770.4;TP391.41

【参考文献】