Recognition Of Diabetic Retinopathy Based On Attentional Mechanism Neural Network

In recent years,convolutional neural network has played an increasingly important role in the medical field and solved many intractable problems in medicine.Application of convolutional neural network in medical image recognition can improve the image recognition rate to a large extent.In the actual diagnosis of diabetic retinopathy,the pathological features are difficult to distinguish with the naked eye,the recognition rate is low,and depends on the clinical experience of ophthalmologists.At present,the use of convolutional neural network to classify diabetic retinopathy can provide doctors with a reliable basis for judgment.In this paper,an approach based on attention convolutional neural network is proposed to classify diabetic retinopathy,and the following studies are carried out mainly from the aspects of network structure:Firstly,to alleviate the gradient vanishing problem,enhances feature reuse is beneficial to feature propagation.However,because there is a direct connection between every two layers in traditional DenseNet,overfitting and large memory consumption problems can occur.To solve these problems,a 2-Densenet model is proposed.The original data set is preprocessed by denoising,image normalization,black edge removal and data enhancement.Then the 2-DenseNet model is built and the model parameters are set.Finally,the preprocessed images are trained and tested,and the experimental results show that the model reduces the parameters in the network,and improves the convergence speed and generalization ability of the network.Secondly,the attention mechanism is introduced into the convolutional neural network,so that the network can automatically obtain the important information in the image feature channel.Previous studies have shown that the addition of SE attention module to ResNet,Inception and Inception-ResNet traditional networks can significantly improve network performance,but the experimental effect is not obvious when the module is introduced into 2-DenseNet.In order to further improve the network performance,the SE module is improved and a new attention mechanism module is proposed.The attention module is embedded into the 2-DenseNet model to guide the network to pay attention to the features such as exudes,thick vessels and microaneurysms in the retinal images.The improved model is used to train and test the preprocessed images,which can further reduce the classification error rate and improve the network performance and classification accuracy.Finally,through a series of comparative experiments,it is verified that the classification recognition rate and network performance of the proposed model for diabetic retinopathy are higher than those of other models.