Research Of Attentional Mechanisms Involved In The Recognition Of Diabetic Retinopathy

Diabetic retinopathy is a complication of diabetes and one of the main causes of vision loss.Early detection and treatment are essential to prevent vision loss.Ophthalmologists diagnose diabetic retinopathy by analyzing fundus images,which is not only inefficient but also labor-intensive and dependent on the experience of ophthalmologists in the field.Deep learning has made significant developments in the field of medical image processing and can be used as an aid for medical professionals.Fundus images,as a major diagnostic reference,show sparse lesions,small differences in adjacent grades,and differences in the same lesions suffered by patients of different genders,ages,and races.Based on this,this paper proposes using the Attention Mechanism,which captures key cues,for grade recognition of diabetic retinopathy.For this point of departure,this paper investigates the performance of the Attention Mechanism in a convolutional neural network,the self-attentive network Vision Transformer,in recognizing the grade of diabetic retinopathy.The research work carried out is as follows.To address the problem of the unbalanced number of classes in the dataset,this paper proposes to use VQ-VAE to reconstruct the images after affine transformation to enrich and balance the dataset.The test results show that the average reconstruction error of the model is0.0001 and the mean structural similarity between the reconstructed image and the original image is 0.967,which proves that the reconstructed image is different from the original image and belongs to the same kind of image,which not only expands the dataset but also increases the Diversity.In the research of attention mechanism in a convolutional neural network for diabetic retinopathy grade recognition,this paper proposes the convolutional channel attention network Res Ne Xt50-SE,Res Ne Xt50-ECA,Res Ne Xt50-SK,convolutional spatial attention network by modifying the network structure and embedding different attention modules with Res Ne Xt50 as the backbone network Res Ne Xt50-Simam,the convolutional hybrid attention network Res Ne Xt50-SPA,Res Ne Xt50-CBAM,Res Ne Xt50-sc SE,Res Ne Xt50-SE+Simam.The nonattentional network Res Ne Xt50 was compared with the convolutional attentional network through experiments.The convolutional attentional network outperformed Res Ne Xt50 in six performance metrics: Precision,Sensitivity,Specificity,F1 Score,Quadratic Weighted Kappa score Qks,Accuracy,and performance against Salt and Pepper Noise,Gaussian noise,and gradient perturbation.Finally,heat maps for each model’s recognition of fundus images were plotted by the Grad-CAM method,and the heat maps showed that the attention network was more effective than the non-attention network Res Ne Xt50 in focusing on fundus images.In the research on the self-attentive mechanism for diabetic retinopathy grade recognition,this paper proposes to use the Vision Transformer,a pure attention network constructed by relying on the self-attentive module,for grade recognition of diabetic retinopathy.A comparison was made between Vi T-B-16,Vi T-B-32,Vi T-L-16,Vi T-L-32 and the convolutional attention network in terms of six performance indexes,namely Precision,Sensitivity,Specificity,F1 Score,quadratic weighted Kappa coefficient Qks,Accuracy and the performance against salt and pepper noise interference and gradient perturbation.Vi T-B-32,Vi T-L-16,Vi T-L-32 outperformed the convolutional attention network,while the performance against Gaussian noise disturbance was weaker than that of the convolutional attention network.Heat maps of the two types of networks recognising fundus images show that the Vision Transformer model focuses on fundus images in a more detailed and thoughtful way.Finally,comparing this paper’s convolutional attention network and Vision Transformer with the work of other researchers,it is found that this paper’s network model outperforms in different degrees on different metrics.A series of experiments and comparisons have shown the superiority and research potential of attentional mechanisms in the recognition of diabetic retinopathy.