Research On Classification Method Of Diabetic Retinopathy Based On Deep Neural Network

Diabetic Retinopathy(DR)is an eye disease caused by diabetes and is the main cause of blindness.Early diagnosis of DR Plays a major role in preventing vision loss and blindness.It is necessary to use efficient and accurate automated screening technology to assist doctors in DR Diagnosis.At present,deep learning method is widely used in the field of DR Classification,and has achieved good results.DR Classification problems based on deep learning can be divided into two categories:binary classification and five-classification.At present,the research mainly focuses on the five classification directions.According to the international standard,DR Lesions are divided into five degrees of severity,which is fuzzy rather than accurate.In medicine,the development of DR Lesions itself is a continuous process.Traditional classification networks regard category labels as accurate and do not take into account the continuity of each lesion grade,which will undoubtedly lose some information.Therefore,many researchers hope to improve the classification effect by using the continuity information of DR Problems.These works mainly fall into two directions.On the one hand,there are efforts to blur labels or increase penalties based on different incorrect predictions based on classification models.These practices have achieved certain results,but the network model always calculates the loss for each sample separately,and cannot explicitly learn from the interrelationship between samples to the internal relationship between categories.On the other hand,some people treat DR Grading as a regression problem entirely,and use regression models instead of classification models to make predictions.In this approach,though,there is modeling of information about the continuity of lesions,the categorical information provided by the data set is not enough to support the pure regression model to achieve particularly good results.In order to make full use of the classification label of the data set,explicitly model the attributes of the continuity of lesions,and give full play to the respective advantages of the two aspects of work.In view of the existing problems,this paper designed the RL-Res Net network,and calculated the classification prediction and regression prediction of samples at the same time.The loss function Re PLoss was designed to model the continuity attributes of lesions explicitly by using regression prediction.Specific work is as follows:1.In order to combine the advantages of classification model and regression model,this paper designs RL-Res Net based on Res Net50 network.The classifier of RL-Res Net network adds one dimension on the basis of the five dimensions,uses the original five dimensions to calculate the classification prediction of samples,and uses the classification prediction to calculate the cross-entropy loss.The added dimension uses the idea of regression learning to calculate the regression prediction of the sample.In order to use regression prediction to train the network to learn the continuity between lesion levels,we designed the relative position loss function Re PLoss.Re PLoss first calculates the prediction distance and label distance between two samples,and then calculates the difference between the two distances as a loss term to constrain the network’s learning of sample distribution.Finally,the cross-entropy loss and Re PLoss are added as the loss function of the whole network to train the network.2.Experiments were carried out on Kaggle’s EYEPACS fundus image dataset and compared with several models with better results in recent years.Our approach achieved 60.27% of ACA and 78.69% of quad-Kappa,using less training data.Moreover,the classification performance of the five categories is more balanced,and the accuracy rate,recall rate and F1-score score are improved compared with the backbone network.It also has certain advantages in DR1 and DR2 samples,which are difficult to classify.In the case of classification errors,the RL-Res Net errors were more subtle,and the overall distribution of the predicted results was closer to that of the label.In addition,the structure parameters designed in this paper are small and simple,which can be easily combined with other DR Classification models.