Research And Application Of Classification Of Diabetic Retinopathy Based On Convolutional Neural Network

In recent years,many scholars have found that convolutional neural networks(CNNs)have achieved good results in many aspects of the medical field.In response to the increasing prevalence of diabetic retinopathy(DR)among the diabetic population in China,we found that using a CNN model to diagnose and classify DR fundus images could help doctors achieve quick and accurate diagnoses,and slow down the growth of the number of patients.In this thesis,we propose a CNN model called Inception ADL,which is based on traditional models and uses transfer learning to obtain initial weight parameters.We have improved the baseline model and optimized its core algorithm.We study the application of this model to the automatic diagnosis and classification of DR fundus images and demonstrate its performance through experiments.To address the difficulty of automatic diagnosis of diabetic retinopathy images and the low accuracy of classification,we propose a new solution: Inception ADL,a CNN-based model for classifying diabetic retinopathy.We use the Eye PACS dataset from the Kaggle competition as the original dataset and use various data preprocessing methods,including binary cutting to remove black background information,normalization to make the light distribution of images uniform,Gaussian filtering to suppress image noise,and CLAHE to enhance image contrast.We also use several data augmentation techniques to expand the dataset and solve the problem of imbalanced image categories.We use the Inception V3 network model as the baseline model and use transfer learning to obtain the initial weights.Since the convolutional layer of the baseline model cannot extract diverse features and the network parameters are large,we introduce dilated convolution to increase the network’s receptive field.We also add fully connected and dropout layers to speed up network convergence and use the Leaky Re LU function after the fully connected layer to extract more effective information.These improvements help improve the performance of the diabetic retinopathy classification model.We input the preprocessed images into the network model and use various techniques to optimize model training during the training process to speed up network training.We then design two sets of experiments for comparative analysis,and the results show that the Inception ADL network model proposed in this thesis outperforms traditional network models in terms of classification accuracy.Through our research,we have achieved accurate classification of diabetic retinopathy fundus images,solving the problem of difficulty in automatic diagnosis and low classification accuracy.