Study On Stage Of Diabetes Based On Deep Learning For Fundus Image

Diabetic Retinopathy(Diabetic Retinopathy,DR)is one of the common chronic c omplications of diabetes and it is a kind of lesion that causes damage to retinal micro vessels.Once a DR occurs,it will seriously affect your eyesight and people with serio us circumstances will cause permanent blindness.Accurate staging of diabetic retinopath y is an important basis for treatment by ophthalmologists.With the development of ar tificial intelligence(Artificial Intelligence,AI),deep learning algorithms have gradually begun to be applied in the medical fields.In this article,deep neural network algorith ms are used in the staging diagnosis of DR aiming to improve the accuracy of DR cl assification,reduce the rate of misdiagnosis and missed diagnosis,so as to reduce the burden on doctors and at the same time to win valuable treatment time for patients.Based on the stage study of diabetic retinopathy,a method of diabetic stages base d on deep learning for fundus image is proposed in this paper.Firstly,the data is pre processed.Then using the classic InceptionV3 network to detect the pre-processed imag es by stages.Then,improving it based on the InceptionV3 model by adding the maxim um pooling layer,convolution layer and batch normalization layer and so on so as to improve the performance of the model by increasing the depth of the model.The first and second transfer learning is carried out on the basis of the classic InceptionV3 mo del.The transfer learning method further proves the effectiveness of the improved mo del compared with the classic InceptionV3 model and transfer learning.The improved model improves the classification accuracy of DR.Finally,the intermediate activation la yer and the heat map of the InceptionV3 network are visualized,aiming to adjust par ameters and optimize the model through visualization.At the same time,TensorBoard i s used to visually display the loss,accuracy,and custom monitoring learning rate indi cators,model structure and other information during through the browser.By monitoringvarious indicators,we can understand,debug,and optimize the network more intuitivel y and conveniently.The experiment uses Kaggle and Messidor two public data sets tov erify the proposed method.Experiments show that in the Kaggle data sets,the classifi cation accuracy is increased from 85% of the original model to 92% of the improved model;in the Messidor data sets,the classification accuracy is in creased from 87%of the original model to 94% of the improved model.In addition to the accuracy rate,the paper also uses the precision rate,the recall rate and the F1-score value to comprehen sively evaluate each category of the model before and after improvement.The experim ent shows that compared with the traditional method,the accuracy rate,the recall rate and F1-score value have been greatly improved.DR can be classified accurately,improve the recognition accuracy of DR,reduce the misdiagnosis rate and the missed diagnosis rate by studying the stages of diabetic retinopathy images.It is an important auxiliary diagnosis and treatment method for opht halmologists and it is of great significance for improving the stage diagnosis of diabet ic retinopathy.