Diabetic Retinopathy Detection And Grading

Diabetic retinopathy(diabetic retinopathy,DR)is one of the most important complications of diabetes.It is extremely important to protect the eyesight of patients to discover the DR at an early stage and perform targeted treatment on it.Based on color fundus images and fluorescein fundus angiography images,this paper studies the segmentation of various physiological structures of fundus images and the detection of multiple types of DR lesions.After the detection of lesions is realized,DR is classified according to the types of lesions in the image,and a high-performance DR lesion detection and grading auxiliary diagnosis system is realized.The main research contents are as follows.(1)A morphological-based blood vessel segmentation method is proposed,and the unique divergence characteristics of blood vessels are studied,and the segmentation of small blood vessels is optimized to realize fast and effective blood vessel segmentation.In order to further optimize the continuity of the blood vessel segmentation results of fluorescein fundus angiography images,a highly continuous blood vessel segmentation algorithm based on Radon transform is also proposed.The overlapping sliding window is used to ensure the continuity of blood vessels,and is based on the Radon transform peak values of different angles of linear structures.The adaptive threshold segmentation of sub-graph blood vessels achieves an accuracy of 95.82% on e-optha.In addition,in the aspect of optic disc segmentation,a method of optic disc positioning based on the results of blood vessel segmentation and template filtering is proposed,and SLIC superpixel segmentation is used to achieve fine segmentation of the optic disc.(2)In the detection of lesions,the detection of microaneurysm is realized based on the Radon transform characteristics of the similar circular structure for the fluorescence contrast image.In addition,SIFT and RANSAC algorithms are used to register fluorescein fundus angiography images at different times,and the detection of neovascularization is achieved through the difference in fluorescence diffusion.Aiming at the color fundus images,an exudation detection method based on mathematical morphology is proposed,and an exudation detection method based on Mahalanobis distance and local linear regression is further proposed.This algorithm has a better detection rate.Finally,based on the SLIC superpixel segmentation and feature difference,the hard and soft penetration classification method is realized.The algorithm in this paper has achieved high detection indexes for all kinds of lesions.(3)In order to further optimize the detection rate of microaneurysm and neovascularization,a joint analysis method of the detection results of the two types of images is proposed.First,the high-precision registration of the two types of images is realized,and then the lesion detection results of the fluorescein fundus angiography image are mapped to the color fundus image,and the false detection area is removed according to the intensity of the red component,and the detection rate based on the single type of image is optimized.In addition,the detection result of MA was further optimized by the detection result of neovascularization.The average precision of images with joint analysis increased by 3.42%,and the F-score value increased by 0.90%.Finally,DR is classified based on the type of lesions in the detection results,and a high-performance DR image lesion detection and grading system is realized.