自动化视网膜血管网络的分形维数定量分析

摘要

目的:研究眼底视网膜血管网络的分形维数自动化测量方法,比较不同分形维数特点.方法:选取20张眼底图片,分别手动分割图像血管和自动分割血管图像计算盒维数、傅里叶分形维数、Hausdorff维数.比较手动分割和自动分割对这三种分形维数的影响.在前面的20张彩色眼底图像中加入高斯噪声和椒盐噪声,比较加入噪声前后对傅里叶分形维数的影响.结果:手工分割和自动分割相比,自动分割血管得出的三种分形维数曲线变化更平稳一点,不像手工分割的分形维数曲线变化忽高忽低,看不出大致的曲线走势和规律.比较来说,自动分割眼底血管图像所得出的分形维数更规律一点.没有加入噪声前对傅里叶分形维数代码进行改造,计算得出的傅里叶分形维数要比原来的数值整体都低且整体趋势更倾向于平稳.结论:自动化视网膜血管网络的分形维数定量研究能够为视网膜病变的分析提供一定帮助.%Objective: To study the automated measuring method of fractal dimension that related to retinal vascular network, compare characteristics of different fractal dimension and the changes of fractal dimension under different pathological conditions. Methods: In our study, a total of 20 fundus image were selected and the retinal blood vessels were segmented manually and automatically. Then, the box dimension, Fourier fractal dimension, Hausdorff dimensions were calculated respectively. The effects of three dimensions after two different segmentation methods were compared. Gaussian, salt and pepper noise was added on the 20 color fundus images and the influence of them on Fourier fractal dimension was tested. Results: Compared to manual segmentation and automatic segmentation, automatic segmentation of vessels curve that drawn three fractal dimension was a little more stable, unlike the manual segmentation curve that can not see the approximate trend and laws of the curve. In comparison, the fractal dimension curve of automatic segmentation fundus image obtained was a little more regular. Whether added Gaussian noise or salt and pepper noise, the fractal dimension obtained to be higher than the original Fourier fractal dimension. In the situation of adding noises simultaneously, the value of Fourier fractal dimension that improved was to be far lower than the value of unmodified Fourier fractal dimension. The curve of multifractal dimension obtained after the wavelet transform was more violent and the waveform was not smooth compared to the original value that obtained from the original images. Contrastively, the curve of multifractal dimension of the original image was in a relatively stable state, almost no major fluctuations. Conclusion: Quantitative study of fractal dimension retinal vascular network could provide some help on retinal image analysis.