The Research On Fundus Image Processing For Aided Diagnosis Of Diabetic Retinopathy

Diabetic retinopathy also known as DR,recognized by WHO(World Health Organization)as one of the three leading eye diseases in the world,is a serious fundus complication in the course of diabetes,and it is also the main reason for blindness and visual decline in the world.There may not be any visual symptoms in the early stage of DR,but the eye has been irreversibly damaged when patients suffer from visual impairment.Therefore,early screening is very important for the treatment of DR.Currently,screening for DR patients is mainly based on the diagnosis and analysis of ophthalmologists for the fundus images of diabetics.It is difficult for the large-scale diabetic patients to achieve regular large-scale screening.With the continuous development of science and technology,computer aided diagnosis is more and more applied to clinical diagnosis and treatment.Fundus images contain a lot of effective information,and image processing technology can be used to analyze the physiological structure and pathological characteristics of fundus images.With the help of computer-aided diagnosis technology,ophthalmologists can achieve large-scale early screening of DR.The main work of this paper is the research of image processing technology in the computer aided diagnosis system for DR.Using the fundus image data obtained by ophthalmology camera,we can achieve the segmentation of retinal vessels,the location and segmentation of optic disc and the detection of the hard exudations which are the obvious features in fundus images of early DR,so as to provide effective help for the aided screening of DR.The main work of the paper includes the following aspects:(1)In the preprocessing phase of retinal blood vessel segmentation,CLAHE(Contrast Limited Adaptive Histogram Equalization)algorithm is used to enhance retinal blood vessel network,especially the tiny blood vessels at the end of vessels.(2)In retinal vessel segmentation process,PST(Phase Stretch Transform)algorithm is used for retinal vessel segmentation.Aiming at the problem of noise on the blood vessel and vascular rupture in the process of PST algorithm,the accurate segmentation of retinal blood vessels is realized by combining different scales of Gaussian filtering and PST algorithm.(3)In the work of optic disc localization and segmentation,this paper proposes a method of optic disk localization and segmentation based on morphological profile analysis.First,the gray morphology opening operation is used to weaken the blood vessels in the fundus image,so as to solve the problem that the optic disk is divided into several blocks;then optic disk is roughly located by adaptive threshold,and by analyzing the contour features in the fundus image optic disk is accurately located;finally,according to the location of the optic disk,ellipse fitting is used to determine the edge of optic disk,and the result is ideal.(4)In the detection of the hard exudations,the deep convolution neural network is used to train the hard exudations classification model.The design of network structure based on the classical LeNet-5 network is suitable for hard exudation detection.In the result show of the detection of hard exudates,in order to provide aided diagnostic result directly,hard exudates probability maps and pseudo color map is calculated.Verified on the open dataset of fundus images,the results of hard exudations are consistent with the doctor’s annotation map,which shows the effectiveness of this method.