Classification Algorithm Of Diabetic Retinopathy Based On Convolutional Neural Network

In recent years,the method of convolutional neural networks has achieved rapid development in the medical industry,solving more and more medically difficult problems.Convolutional Neural Networks(CNN)and modern medical equipment are combined to provide a reliable basis for the diagnosis of diseases through a series of convolutional neural network technologies.This has become the future trend of medical development.The diagnosis of diabetic retina has mostly relied on the experience of doctors,and the number of ophthalmologists is far from meeting the demand.The lack of medical conditions makes it impossible for patients' diseases to be diagnosed and treated in time,and even leads to irreversible blindness.Therefore,the use of Convolutional Neural Networks for automatic image classification and classification has an important role in the prevention and treatment of diabetic retinopathy.In this paper,for the purpose of obtaining better detection effect of diabetic retina,the improved VGG network structure and the design of convolutional neural network structure XNet are integrated,and a set of B/S structure diabetic retina image detection system is constructed.The main research of this article is as follows:(1)Describes the theoretical knowledge required for diabetic retina detection,including Convolutional Neural Networks and classic network models,supervised learning algorithms such as transfer learning,integrated learning,and system design related knowledge such as deep learning framework TensorFlow,front-end framework Boostrap,and back-end framework ThinkPHP.In order to determine whether the diabetic retinopathy has occurred,the retinal image data set provided by the Kaggle competition abroad is used for two classifications,and an automatic retinal image diagnosis system based on Bagging integrated learning is designed.This algorithm integrates two improved VGG convolutional neural networks,and adds Leaky-ReLU as the activation function.In one of the improved models,each convolutional layer uses Dropout regularization,the overall network structure is moderate,and GPU can be used to speed up,using the voting method to calculate the final classification results.Finally,for different algorithms,the accuracy of different iterations is compared.The experimental results show that the accuracy of classification can be improved after integrating the two networks,and the accuracy is increased by 2%to 82%.(2)Aiming at the need to detect the degree of diabetic retinopathy,using the retinal data set provided by the Hainan Province Eye Hospital,a Convolutional Neural Network XNet for retinal image lesion grading is designed,which has a moderate structural complexity.For computer hardware.The requirements are low.The network parameters are adaptively adjusted according to the training samples.After preprocessing such as data cleaning,normalization,and data amplification on the data set images,a part of the data is retained as the verification set,and the classic convolutional neural network LeNet and the accuracy and the number of iterations of the Inception-V3 network structure are compared.The experimental results show that the structure of XNet network is better than LeNet and Inception-V3,and the accuracy rate can reach 91%.(3)In response to the needs of assisting doctors in diagnosing the degree of diabetic retinopathy,a fully automated diabetic retina detection system was completed using the XNet network as the basis.The B/S structure was used for website architecture,combined with the front-end framework Boostrap framework and the back-end framework ThinkPHP framework.The entire system is divided into several models:login registration module,image upload module,image detection module,and history recording module to design and implement visual website interface design of diabetic retina detection system based on Convolutional Neural Network..