A Study On Semantic Segmentation Of Fundus Image For Anatomical Structures

The study of fundus image analysis and processing technology has important clinical guiding significance for the diagnosis of eye diseases and other human body diseases.However,the vessels and optic disc are important parts of the anatomy of the fundus,which have different features and complex structures,making it difficult to achieve the segmentation of the two types of anatomical structures at the same time.In this paper,we have adopted the integrated learning algorithm to realize the segmentation of blood vessels,and applied Bayesian inference to realize the segmentation of optic disc.On this basis,the semantic segmentation of the two anatomical structures has been realized with decision tree for the first time.Specific work is as follows:Firstly,a retinal vascular segmentation method has been proposed based on AdaBoost.First,a series of preprocessing performed on the fundus image enhances the contrast between the blood vessels and the background while removing the noise.Then,the 7-dimensional vessel characteristics such as G-channel reversal gray value,B-COSFIRE,Line detector and 2D Gabor wavelet are extracted.These features and database annotated by experts have been fed to the AdaBoost classifier training model.The DRIVE and STARE databases are used to achieve accurate segmentation of blood vessels with the trained model.Secondly,a method of segmentation of optic discs has been proposed based on rules,which applying a coarse to fine strategy.First,the region of interest is detected,and the convex hull is calculated using the color-chanced Harris,which resulting in a rough optic disc area.Then,the prior probability distribution is computed based on the approximate optic disc area and the super-pixel segmentation result.At the same time,the probability of observation of each pixel is calculated based on the convex hull statistical color histogram,and the posterior probability distribution is obtained according to the Bayesian inference.Finally,the image has been binarized and the optic disc has been precisely segmented in the open database RIM-ONE using standard Hough transform circular detection.Thirdly,a semantic segmentation method of retinal image anatomy has been proposed based on decision tree.With the same idea of blood vessel segmentation method,the fundus images are preprocessed firstly.The gray value of G-channel is reversed,B-COSFIRE and Gabor wavelet are used as the features of the blood vessels.And the posterior probability of the above Bayesian inference and another simple saliency result are used as the characteristics of optic disc.With the help of ophthalmologists in Shengjing Hospital,the database for experiments has been created.And the decision tree classifier has been chosen to realize the semantic segmentation of three kinds of blood vessel,optic disc and background.Finally,this paper has developed the GUI interface in the GUIDE environment provided by MATLAB.According to the above methods,the interface designing and component programming are realized.At last,the retinal vascular segmentation system based on AdaBoost,the retinal optic disc segmentation based on the rules,the semantic segmentation system of fundus anatomical structure based on decision tree have been realized respectively.