Research On Detection Of Diabetic Retinopathy Based On Deep Learning

Diabetic retinopathy(DR)is a common,long-term complication that can cause significant visual damage if not properly diagnosed and managed.Due to the diversity and complexity of the condition,manual DR detection is time-consuming and laborious with great uncertainty.To address the above problems,this thesis adopts a deep learning approach to intelligently identify and classify DR images by feature extraction to complete the graded detection of DR,enhance diagnostic accuracy and improve efficiency.The main work of this thesis includes:Firstly,the open source Kaggle dataset is selected,and the dataset is standardized,normalized,Gamma adjusted,CLAHE histogram equalized,and data enhanced for the problems of large differences in image quality and unbalanced sample size of different types.The new dataset with the required image quality and balanced sample data is obtained to provide sufficient pavement for the future model training.Next,the data set is trained from both convolutional neural network and Transformer model,and the experimental results are compared and analyzed by the evaluation criteria of accuracy,precision,sensitivity and specificity.For convolutional neural networks,AlexNet,MobileNet,ShuffuleNet network models and their variants are used for comparison experiments and analysis.The models were ranked in descending order of accuracy as MobileNet v3,ShuffuleNet,MobileNet v2,AlexNet v1,and AlexNet v2.Among them,the MobileNet v3 model with the best overall effect achieved an accuracy of 87.54%.For the Transformer model,the Vision Transormer model and the Swin Transformer model were used to compare the experiments on whether to use migration learning or not,and the results were compared with those of the convolutional neural network.The best model is the Swin Transformer model using migration learning,with an accuracy of 89.54%.The results illustrate the superiority of the Transformer model over the convolutional neural network in the field of diabetic retinopathy detection for recognition and the importance of migration learning for the Transformer structural model.In addition,this thesis visualizes the classification effect of each network by confusion matrix and finds that the model in this experiment is the best in classifying samples from category 0,while misclassification occurs to varying degrees for both category 1 and category 2.The reason is that the key points of "bleeding point" and "microaneurysm" are too similar between category 1 and category 2.The model could not distinguish them well.Finally,the optimal model Swin Transformer was used as the main framework,and the TensorFlow Lite framework and Android Studio platform were used to design an automatic diabetic retinopathy detection app,which can diagnose and produce results in real time,and to a certain extent,alleviate the problem of medical resource constraints.