Image Recognition Of Diabetic Retinopathy Based On Efficient Neural Network

Diabetic Retinopathy referred to as DR,is a chronic complication caused by diabetes.It is called Diabetic Retinopathy because the diabetic patients with hyperglycemic in a long-term environment,the hemorrhage,exudate,microhemangioma and other lesions will slowly appear on the retina,eventually causes the patient’s vision to decline and even blindness.At present,the main method of prevention still relies on the regular photography of fundus images for screening.Due to the limitation of medical resources,the large patient set can only rely on a few professional technicians and clinicians to take images and diagnose.This process is very time-consuming and laborious,which not only delays the treatment of patients,but also leads to the increasing rate of misdiagnosis.With the rapid development of deep learning technology and computer performance,it is possible to use computer aided diagnosis.Due to the limited energy of manual analysis and diagnosis,a large amount of effective information in fundus images is not used.The method of deep learning is very powerful in feature extraction,so this paper proposes a method of automatic recognition of sugar mesh image based on deep learning algorithm.First,we used the residual network model and the Inception network model to identify and classify the fundus image of DR,and then compared with the algorithm base on the traditional convolutional neural network use the fundus image of DR.Experiments show that extending the network in any dimension,the performance of the extended network is better than that of the traditional convolutional neural network.However,when the model is extended to a certain extent,the phenomenon of degradation occurs and the amount of parameters becomes larger and larger.Then,we propose a DR detection method based on the inception＿resnet V2 model.This method combines the extension of the two dimensions of network depth and width,and proves that the performance of the model obtained by two-dimensional scaling of depth and width.which is better than the one-dimensional scaling.It also solves the phenomenon of network degradation,but the accuracy of classification has not been greatly improved.Finally,we propose a method for detecting fundus images of DR based on EfficientNet network.This method scales the model from the three dimensions: the depth,the width and the resolution of the image.and the network model performs best when the three dimensions reach a balance.The accuracy of the most basic version of this method is higher than that of the other three network models,and the classification accuracy of the best version reaches 88.57%.Not only the accuracy is high,but the performance in all aspects is better than other models,and the sensitivity to the lesion is also high.In order to verify the performance of the model,the experimental data used in this article is a two-year public data set provided on the kaggle platform.A large number of experimental comparisons were made to obtain the best network version Efficient B5 and the optimal parameters.The experimental results show that the deep learning network model used in this paper can effectively detect and classify the lesions in the image of sugar retinopathy.