眼底图像中视盘的定位与分割方法研究

发布时间：2018-07-25 07:51

【摘要】：眼底是观察全身疾病的“窗口”,高血压、高血脂、肾炎、糖尿病、中枢神经系统等许多疾病均会引发眼底病变,因此基于眼底图像的疾病早期筛查和检测越来越受到重视。视盘是眼底图像的一个重要特征,对其进行准确、快速地定位与分割对利用眼底图像进行疾病辅助诊断具有重要意义。在兼顾定位与分割准确性的同时,本文所研究的算法着重关注计算速度以满足实际系统对实时性的要求,为此主要从感兴趣区域确定、视盘候选区选取和血管擦除三方面进行了算法改进。具体研究内容如下:(1)预处理方面,在眼底图像的红色分量中使用大津阈值分割法生成感兴趣区域掩膜图像,使用双线性插值法将感兴趣区域的边长缩放至540个像素;(2)血管分割方面,在眼底图像的绿色分量上对比四种对比度增强算法,选取了速度快且效果优秀的CLAHE(Contrast Limited Adaptive Histogram Equalization)方法。对比了基于自适应直方图均衡及形态学底帽变换的两种血管增强算法,选择了耗时较短的后者。在增强结果上使用大津阈值分割得到血管图像;(3)视盘定位方面,基于血管分割结果对感兴趣区域进行取半操作以减少后续计算量。依据视盘的亮度特性,使用DLC(Directional Local Contrast)算法获取眼底图像中的亮斑区域作为视盘候选区,随后依据视盘的血管特性选取出血管分支数最多的候选区为视盘区域,取该区域重心坐标为视盘中心定位结果;(4)视盘分割方面,结合形态学方法与插值方法得到了快速有效的血管擦除方法,在血管擦除结果上使用Canny算子进行边缘检测,使用Hough变换得到视盘分割结果。最后完成了对包含不同眼底病变的MESSIDOR眼底图像库(1200张眼底图像)的测试。结果表明,本文研究的算法具有较好的视盘定位效果,得到了99%的准确率,平均用时0.777s。在定位正确的结果上进一步得到了较好的分割效果,平均耗时0.295s。因此,可以说本文研究的视盘定位与分割算法具有复杂度低、速度快,准确率高等优点,相对其它方法更符合实际系统实时处理要求。
[Abstract]:Fundus is the "window" to observe systemic diseases. Many diseases such as hypertension, hyperlipidemia, nephritis, diabetes, central nervous system and so on can cause fundus diseases. Therefore, the early screening and detection of diseases based on fundus image are paid more and more attention. Optical disc is an important feature of fundus image. It is very important to locate and segment the image accurately and quickly for the diagnosis of disease using fundus image. While taking into account the accuracy of location and segmentation, the algorithm studied in this paper focuses on the calculation speed to meet the real-time requirements of the actual system, so it is mainly determined from the region of interest. The algorithm is improved in three aspects: the selection of disc candidate area and vascular erasure. The specific research contents are as follows: (1) in the aspect of preprocessing, the region of interest mask image is generated by using Otsu threshold segmentation method in the red component of the fundus image. Bilinear interpolation is used to scale the edge length of the region of interest to 540 pixels. (2) in the aspect of blood vessel segmentation, four contrast enhancement algorithms are compared in the green component of the fundus image, and the fast and excellent CLAHE (Contrast Limited Adaptive Histogram Equalization) method is selected. Two kinds of vascular enhancement algorithms based on adaptive histogram equalization and morphological bottom hat transform are compared. In contrast to the enhancement results, we use the Otsu threshold to segment the vascular images. (3) in the aspect of visual disk location, the region of interest is half-operated based on the segmentation results to reduce the amount of subsequent computation. According to the luminance characteristics of the disc, the bright spot area in the fundus image is obtained by DLC (Directional Local Contrast) algorithm as the candidate area of the disc, and then the candidate region with the largest number of vascular branches is selected as the visual disk area according to the vascular characteristics of the disc. The center of gravity of the region is taken as the center of the visual disk. (4) in the field of disk segmentation, a fast and effective vascular erasure method is obtained by combining morphological method with interpolation method, and the edge detection is performed by using Canny operator on the result of vascular erasure. The Hough transform is used to get the result of the disk segmentation. Finally, the MESSIDOR fundus image library (1200 fundus images) with different fundus lesions was tested. The results show that the algorithm presented in this paper has a good location effect, and the accuracy is 99%. The average time used is 0.777s. A better segmentation effect was obtained on the basis of correct localization results, with an average time of 0.295 s. Therefore, it can be said that the video disk localization and segmentation algorithm studied in this paper has the advantages of low complexity, high speed, high accuracy and so on. Compared with other methods, it is more suitable for real time processing of real time system.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】