Classification Of Diabetic Retinopathy Based On Dual-stream Fusion Network And Attention Mechanism

Diabetic retinopathy(DR)is one of the most serious complications of diabetes,which is a kind of fundus lesion with specific changes.At present,the diagnosis of retinopathy mainly relies on artificial screening.Clinically,due to inadequate medical resources,many patients are not treated in time,resulting in severe lesions and even blindness.In addition,only relying on the manual diagnosis of doctors,the results are subjective and inaccurate.Due to the variety and different morphology of DR lesions,automatic classification of fundus images in mass screening can greatly save clinicians’ diagnosis time.With the wide application of artificial intelligence,deep learning technology has shown excellent performance in the field of medical image processing.Therefore,the development of a computer-assisted diagnosis and treatment system to help clinicians complete the early diagnosis of retinopathy can effectively reduce the visual damage caused by retinopathy.In this paper,we combined the graph convolutional network and attention mechanism to study the retinal image classification task.The specific research contents are as follows:(1)In order to solve the problems of uneven distribution of various categories of samples in the dataset,unclear image features and different image size and brightness,this paper uses a variety of preprocessing methods such as image clipping,brightness normalization and border removal to enhance image features.In addition,rotation,translation and other operations are used in the experiment to increase the number of images to reduce the impact of uneven sample distribution.(2)To solve the problem of complex retinal pathological features,this paper proposes a retinopathy image classification method based on dual-stream fusion network and attention mechanism.The network consists of two modules: CNN and GCN,and uses semi-supervised classification to improve the generalization ability of the network.CNN and GCN were used to extract global features and spatial location features of fundus images respectively,and attention mechanism was introduced to enhance the adaptability of GCN to graph.Finally,the features extracted by the two modules are fused and input into the classifier for classification.(3)To verify the effectiveness of the network,we conducted comparative experiments and ablation experiments.We adopted confusion matrix,accuracy,precision,recall and kappa score as evaluation indexes.With the increase of the labeling rates,the classification accuracy is higher.Particularly,when the labeling rate is set to 100%,the classification accuracy of GACNN reaches 93.35%.Compared with Dense Net121,the accuracy rate is improved by 6.24%.Ablation experiments show that semi-supervised classification based on attention mechanism can effectively improve the classification performance of the model,and achieve good results in classification indexes such as accuracy and recall.The proposed method provides a feasible classification scheme for fundus images and effectively reduces the human resources required for screening.