Intelligent Diagnosis Of Diabetic Retinopathy Based On Fundus Image

The diabetic retinopathy is one of the main factors leading to irreversible blindness.Early screening and timely diagnosis and treatment are of great significance.The diagnosis of diabetic retinopathy is to judge the degree of diabetic retinopathy of the patient based on the color fundus image taken by fundus camera.The proposed approach in this paper combines the global features of the whole retina image with the local features of different types of lesions to achieve disease intelligent diagnosis.The main research work of this paper are as follows:A disease diagnosis method based on residual-bilinear network is proposed,which can solve the problem that the difference between diabetic retinopathy images with different severity is very small and the feature extraction is difficult.By introducing bilinear features,we design a residual-bilinear convolutional neural network to extract the rich semantic information of the fundus image,so that the network can pay attention to discriminative areas.The traditional residual block is improved to avoid the loss of some important feature information,and the features between different channels are combined by bilinear pooling to obtain high-level semantic features.In addition,for the problem that some small lesions can not be well expressed on the feature map of deep layer,we proposed a multi-level feature fusion method to get context information from different levels in the network.Compare and analyse on the EyePACS dataset,the proposed method improves in sensitivity and Kappa.To solve the problem that features of different types of lesion cannot be simultaneously extracted under a uniform algorithm framework in the current mainstream lesion detection algorithm,and in order to avoid the small lesion area being ignored,a lesion classification method based on patch-wise fundus image is proposed,which can provide features of lesion for disease diagnosis.The multi-scale feature fusion method is proposed for the problem of large difference in the morphology of different lesion regions in fundus images.This paper introduces atrous spatital pyramid pooling to extract features of lesion in different regions of the fundus image.The experimental results prove that different lesion areas can be well identified on the IDRID database.The proposed approach achieves 84.54% of recall rate and 79.97% of precosion rate of microaneurysms,and 90.84% of recall rate and 95.27% of precosion rate of exudate.A disease diagnosis method based on lesion features fusion is proposed,which can solve the problem that adjacent severity images are easily confused,especially normal and mild images in diabetic retinopathy.To provide prior knowledge of the lesion area,the suspicious image patches are extracted by image processing.In this paper,the whole fundus image and the corresponding suspicious image patch are input into the disease diagnosis network and the lesion classification network respectively,tuning training based on the trained network model.On the EyePACS database,the proposesd approach reaches 89.5% of sensitivity,86% of specificity rate,and 0.846 of Kappa.In summary,this paper proposes diabetic retinopathy diagnosis methods based on residual-bilinear network and lesion features fusion,which provides a new idea for disease diagnosis research.The proposed two methods are verified on the EyePACS dataset,and the experimental results shows the effectiveness of proposed algorithms.