Research On Diabetic Retinopathy Intelligent Diagnostic Method Based On Deep Learning

Diabetic retinopathy(DR)has a higher risk of blindness,which has a great impact on people’s production and life.Intelligent diagnostic methods can alleviate the problem of insufficient medical resources under the demand of large-scale DR screening,which has very important practical significance.At present,deep learning has become the mainstream method in the field of medical image analysis.Many DR five-stage assisted diagnosis methods based on deep learning technology have been proposed continuously and achieved good results.In general,the ideal results are mainly attributed to the use of large amounts of annotated data,and the design of feature extractors and classifiers in the model is also crucial.However,the existing models still face the following challenges: when constructing the network structure,the simple neural network structure lacks the ability to encode and classify the features of lesions;In the training process,due to the existing open source data sets of fundus images,such as device differences and regional differences of diseases,the classification results of the model will be inaccurate.As a result of these problems,the DR diagnostic model for large-scale open source data training is not suitable for some medical institutions with insufficient label data or different countries and regions.In order to reduce the dependence of the model on the amount of label training data while minimizing the accuracy of the model,this paper proposes a DR five-stage intelligent diagnosis network which can also have high reliability under the condition of a small amount of data training.In addition,a method that can use a large amount of unlabeled data to assist model training is proposed.The specific work contents of this paper are as follows:Firstly,based on the deep residual network model,the process of image feature extraction and classification was improved to solve the problems in the model structure,and a DR five-stage diagnosis model(BiRAD-Net)was proposed.In the feature extraction stage,in order to optimize the feature quality,the bybrid attention mechanism was introduced to suppress the noise,and the feature grade decision network module was designed to enhance the coding ability of the model.In the process of classification,a double-branch classifier and its corresponding loss are designed to enhance the accuracy of classification.The experimental results show that BiRAD-Net can improve the recognition ability of each stage of DR while reducing the amount of data required by the model.Second,an improved DR staging network training algorithm(SimCLR-DR)is proposed to solve the problems in the training process.The algorithm consists of two parts: the pretext and the downstream task.In the pretext task,the encoder is pre-trained by using a large amount of unlabeled fundus image data combined with the contrastive self-supervised learning method.In the downstream task,a small amount of labeled data is used to fine-tune the pre-trained encoder and classifier.The experimental results show that the SimCLR-DR algorithm can improve the classification performance of the model to some extent and make it have better generalization ability under the condition of less training data.