The Fusion Characteristics Of Diabetic Retinopathy Eyes Image Recognition Method

Diabetic Retinopathy(DR)is a retinal disease caused by diabetes.The long-term hyperglycemic environment causes damage to the vascular wall,resulting in microangioma,exudate,hemorrhage and other lesions on the retina.DR is one of the reason of blinding because the lesion covers the retina.Currently in clinical practice,ophthalmologists usually use the fundus camera to image the patient's retina,by analyzing and diagnosing abnormal lesions of the retina.However,due to the variety and characteristics of diabetic retinopathy,which makes ophthalmologists difficult to diagnose.With the rapid development of medical image processing technology,computer-aided diagnosis systems have gradually become an important method for doctors to diagnose patients.At present,the main methods for classifying fundus images of diabetic retinopathy are based on the detection method of fundus image lesion features and the method based on the overall feature extraction and classification of fundus image.The method of the lesion's feature extraction requires manual design of feature extraction methods for different lesions,and then do the classify of a classifiers.Because of the variety of lesions and irregular distribution of diabetic retinopathy,the classification of the disease is less efficient and less adaptable.Convolutional Neural Networks(CNN)has the characteristics of combining feature extraction and classifier classification,be able to learn characteristics by itself,using convolutional neural network to classify fundus images of diabetic retinopathy can avoid the cumbersome feature extraction work.In this paper,based on the traditional convolutional neural network model,a fundus model for diabetic retinopathy can be used to classify the disease course.The research content of this paper consists of two parts.On the one part,there will compare the processing effects of different model structures.This part introduces a new feature fusion model based on binocular.It is proved by experiments that this model used to classify the fundus image of diabetic retinopathy will be more effective.The experimental data is a public data set provided by Kaggle.The optimal classification accuracy of the model on this data is 77.54%,which is improved compared with other models.On the other part,it will analyze the optimal learning rate and loss function under different evaluation criteria to adjusting the best loss function and learning rate of the model.Finally,this paper visualizes the feature layer of the convolutional neural network,and verifies that the convolutional neural network is gradually integrated into the highlevel abstract information from the specific features of the lower layer.By analyzes the misclassified samples,this paper found the two main causes which resulting in misclassfying: not obvious feature of microangioma and image artifacts caused by camera artifacts.