The Application Of Convolutional Neural Network In The Classification Of Fundus Images Of Sugar Net Diseases

Diabetic Retinopathy(DR)is one of the major blinding diseases.The high level of glucose in the body results in vascular wall damage for diabetics.The microaneurysm,exudation and hemorrhage lesions will appear in retina.In clinical examination,the most commonly used method is to observe the retina at the back of the eye by fundus camera.Ophthalmologists detect the abnormal regions in the fundus image to diagnose DR.However,it is hard to distinguish DR lesions,because they have many kinds and complex structures.Especially,in the large scale screening,computer aided diagnosis system can help ophthalmologists to diagnose quickly and effectively.At present,the common DR fundus image classification have two methods,one is based on detecting local lesions and the other one is based on full image features.Because these traditional methods need manually extract features,they have low efficiency and poor adaptability.Convolutional neural network(CNN)combines the feature extraction and classification process.It is can automatic learn features within the training.CNN is applied to classify the DR fundus image.For address the problem of insufficient fundus images,this article apply CNN and transfer learning to classify fundus image.This article main contains two aspects.One is feature transfer learning which use many pre-trained models to extract features from fundus image,then these features are applied to train a support vector machine classifier.The other is fine-tuning,we first augment the amount of fundus image and apply many pre-trained models to initialize the parameters of network architecture of fundus image classification,and then train the classier on fundus images.This article conducts many experiments on DR1,Messidor and our dataset.All experiments are based on five-fold cross validation.The best classification accuracy is94.52% on DR1 dataset,the best classification accuracy is 92.01% on Messidor dataset and the best classification accuracy is 97.93%.The results verify the availability of CNN and transfer learning on DR fundus image classification.Moreover,this article analyze the networks features visualization,and find the lower architecture of CNN extract the specific feature,the high architecture extract the abstractedfeature.This article analyze the image which error classification and find the reason is come from these aspects,such as the quality,quantity and variety of fundus images.At the end of this article,we implement a DR fundus image classification prototypical system which can help ophthalmologists diagnose DR.