Fundus Image Segmentation And Auxiliary Diagnosis Based On Deep Learning

Glaucoma is an eye disease with high blindness rate in the world.Early glaucoma has no obvious symptoms.Up to the time of discovery of visual impairment and visual field loss,patients often reach the advanced stage of glaucoma,and the course of disease is irreversible.Clinically,an ophthalmologist performs a fundus examination of a patient using a color fundus image.However,the current clinical diagnosis of ophthalmic diseases heavily relies on the medical experience of ophthalmologists,resulting in a large difference in examination results.In large-scale fundus examinations,doctors spend a lot of time and energy,which leads to misdiagnosis and missed diagnosis.Therefore,this article uses computer-aided diagnostic techniques to assist ophthalmologists in large-scale fundus examinations by combining image processing and deep learning techniques.In the fundus image,glaucoma disease often occurs as optic atrophy and optic disc linear hemorrhage,in other words,it may change optic disc,optic cup and retinal vessels structure.Therefore,the accurate segmentation of optic disc,optic cup and retinal vessels and the measurement of related parameters are of great significance for the preliminary screening and late diagnosis of glaucoma.Aiming at the problems of low accuracy,low sensitivity and incorrect recognition of the segmentation area in existing segmentation algorithms,this paper designs deep neural network models based on the basic theory of deep learning to segment retinal vessels,optic disc and cup,and uses the segmentation results to calculate clinical parameters to assist ophthalmologists in the preliminary screening of glaucoma diseases.The main work of this article includes:Firstly,aiming at the problem that the existing blood vessel segmentation algorithm has low accuracy and low sensitivity for small blood vessel segmentation in fundus images,an improved segmentation algorithm for fundus image of U-shaped network is proposed.The algorithm utilizes the idea of residual network by changing the traditional convolutional layer serial connection method to the residual mapping.It also optimizes the network by adding batch normalization and PReLU activation function between the convolution layers.The algorithm was tested on the DRIVE and CHASE_DB1 fundus databases,with an average increase of 2.47%,0.21%,and 0.35% in accuracy,sensitivity,and AUC compared to the best algorithms in the literature.Secondly,aiming at the problem that the existing optic disc and cup segmentation algorithm has low precision of disc and cup segmentation and misidentification of the segmentation region,a deep learning method is proposed to segment the disc and cup structure of the fundus.In this algorithm,a full convolution neural network is designed and a fully connected conditional random field is used as post-processing to segment the disc and cup area of the fundus image.The experimental results show that the algorithm improves the recognition of the edge of the optic disc region and improves the accuracy of segmentation.Finally,in order to verify the practicability of the proposed algorithm,this paper carried out experiments on 50 clinical color fundus images collected from the hospital to achieve preliminary screening of glaucoma diseases.Using the network model proposed in this paper to segment the retinal vessels,optic disc and optic cup structure,calculate the clinical parameters such as cup-to-disk ratio,nerve retinal marginal ratio and blood vessel ratio to realize preliminary screen glaucoma disease.Experiments have shown that the accuracy of identifying suspected glaucoma diseases can reach 96%.