| At present,the sheer variety of highly processed foods confuses the definition of healthy food.More and more people suffer from abnormal body metabolism due to the long-term intake of unhealthy ingredients,and the blood sugar is at a peak and drops slowly,which in turn induces diabetes.The disease causes retinopathy and,in severe cases,blindness.Therefore,it is necessary to regularly check the retinal condition,and to diagnose and treat the lesions in time.At present,the main method for diagnosing lesions is manual screening by doctors.This process is time-consuming and laborintensive.Coupled with the lack of medical resources in some areas,misdiagnosis may occur.Therefore,the use of computer technology to assist doctors in the diagnosis of diabetic retinopathy is of great significance.Because deep learning technology can automatically learn based on massive data features,it is widely used in image processing tasks because of its unique advantage of capturing important information of data by using its feature extraction ability.Based on this,this paper conducts research on the classification of diabetic retinopathy.The specific contents are as follows:（1）Model design: Image classification in the medical field is biased towards finegrained image classification,that is,the differences between classes are not obvious.Using traditional stacked network layers to obtain more detailed features of the classification network will lead to partial image edges or important lesion features.lost.Aiming at this problem,this paper starts from the Res Net residual network as the basic structure.In order not to reduce the feature extraction due to channel compression,and to maintain the balance between the parameters and channel features,and reduce the amount of calculation,a volume based on channel expansion is designed.The product module,referred to as CDcov Block,replaces the original partial convolution structure.Through point-by-point convolution,the number of channels is expanded to twice the original feature map to accommodate more features.Channel attention is introduced in this module to solve the problem that the network depth has been unable to bring more accuracy gains.This module is used as the backbone module of the classification model CD＿Net designed in this paper.Aiming at the problem of small receptive field of convolution operation and poor ability to capture long-distance information correlation,a self-attention mechanism Non-Local module is added to the last two layers of the network to improve the detail perception ability of the network.Based on the above description,a classification model CDN＿Net is designed.（2）This paper uses the public data set EYEPACS.Due to the non-uniform brightness,contrast,size and data distribution,the learning ability of the model will be directly affected during model training.In this paper,data preprocessing is performed before the model training,including normalizing the brightness of the fundus image using the gray-scale world algorithm,using the binary method to cut the black border of the fundus image,and adjusting the pixels to 512*512,which is suitable for the model in this paper.Learn,use Gaussian filtering to process noise in images and amplify features.The data quality of the input model is unified through the above normalization operation.Then,the sample size is increased by horizontal offset,flip,vertical offset,rotation transformation,etc.to ensure a balanced sample size.（3）Model training and verification: In order to fully prove the rationality of the classification model CDN＿Net designed in this paper,three groups of experiments are set up for verification.Under the same dataset and experimental environment,the experimental results were analyzed using different evaluation criteria,and multiple comparison and ablation experiments were performed.The final sensitivity of the CDN＿Net model on the test set is 92.3%,the specificity is 94.9%,and the accuracy is as high as 96.1%.On the one hand,it proves the rationality and effectiveness of the design of CDN＿Net,and on the other hand,it proves the superiority of this method compared with the traditional network in the classification of diabetic retinopathy. |
PDF Full Text Request