Diabetic Retinopathy Detection Based On Data-enhanced Detection

Diabetic retinopathy(DR)is an eye symptom caused by diabetes,a systemic disease.Patients with diabetes for longer periods of time are often plagued by this eye symptom.The detection and severity grading of diabetic retinopathy is the key to the treatment of diabetic retinopathy.In recent years,convolution neural network(CNN)has achieved remarkable performance in many different visual classification tasks.Because the model of general deep convolution neural network is complex and has a lot of parameters,it needs a lot of labeled train data to get better outputs.However,many applications in reality are difficult to obtain a large number of labeled sample data,so it is of great significance to do model enhancement on a small set of training data.This paper focuses on the research of model integration technology and data enhancement technology,and improves the generalization ability of the final classification model by applying various types of data and integrating multiple learning models.Then the purpose of fast and accurate classification of fundus pictures of different quality was achieved.The main findings are as follows:(1)This paper presents a screening method for diabetic retinopathy based on integrated learning.Firstly,the color histogram equalization method is used to enhance the fundus picture information,and the data expansion technology is further used to improve the performance of the neural network model with different depth of architecture,and finally,the proposed neural network model is used.A new ensemble learning strategy is proposed to solve the problem of lacking training data and complex data boundary,which makes the final classifier more accurate,more stable and more applicable.The results of binary classification of color fundus images show that this method can significantly improve the accuracy of diabetic retinopathy screening.(2)A method for screening diabetic retinopathy based on data augmentation was proposed.In view of the characteristics of fundus images,this paper combines the techniques of generating anti-network and super-resolution to selectively expand the original input data,so as to generate more training samples with detailed information.In addition,in order to solve the problem of different data distribution caused by the inconsistency of fundus images,the idea of style migration is further introduced in this paper,so other fundus images can be transferred to expand existing data set.The experimental results show that the staging framework of diabetic retinopathy based on data augmentation can partly solve the problem of small input data set and uneven distribution of different types of image samples,thus effectively saving the cost of manual annotation.The classification accuracy and kappa coefficient are improved.