Research On Classification Of Diabetic Retinopathy Based On Convolutional Neural Network

Diabetic retinopathy is the leading cause of blindness.Timely detection and diagnosis can greatly reduce the risk of blindness.It is of great significance to study efficient and accurate classification algorithms for the severity of diabetic retinopathy.However,traditional classification models still have the problem that the feature information is not fully extracted and underutilized.Therefore,the multi-scale deep convolutional neural network that integrates attention is studied in this thesis,using the attention mechanism to select high-correlation features and merge the deep and shallow features to achieve efficient intelligent diagnosis of diseases.The main work is as follows:To deal with the single feature problem existing in the existing convolutional neural network when processing image classification,a Multi-scale Deep Convolutional Neural Network(MDCNN)was proposed in this thesis.The model is based on ResNext,a variant structure of Residual Network(ResNet),in which the idea of grouping convolution of aggregated residual network was used to broaden the network dimension to extract diverse features and reduce the computational load of the model.At the same time,the methods of multiresolution image input and multi-layer feature fusion were used to improve the feature extraction ability of the network.Experiments on EYEPACS 2015 of the kaggle competition prove that the multi-scale method enables the network to extract more diverse features,which can solve the problem of single feature in image classification.Aiming at the difficulty of classification due to the small difference between different classes of fundus images,an attention mechanism was introduced in the multi-scale network model and a Multi-scale Attention Deep Convolutional Neural Network(MADCNN)was proposed.Firstly,a variety of pre-processing methods were utilized to process image data which can eliminate the noise information generated during the acquisition process and amplify the negative sample data.Secondly,multi-dimensional parallel convolution kernels were used to extract the features of the same feature map during feature extraction.Then,the relationship between feature channel and classes were learned.Finally,featureswith high correlation were selected for model learning.By conducting experiments on APTOS 2019,a public dataset of the kaggle competition,the multi-scale network with attention mechanism can obtain a better classification effect than the classic network.In addition,in order to provide doctors with more diagnostic information,heat visualization was used in this thesis to represent the feature areas of lesions in the image by calculating the gradient relations between feature areas and classes,so as to better assist doctors in diagnosis.In summary,a multi-scale attention deep convolutional neural network was proposed for the shortcomings of traditional classification models,which can efficiently and accurately classify diabetic retinopathy.And it solves the problem that feature information is not fully extracted and underutilized.