AlexViT: Novel Diabetic Retinopathy Image Classification

In this study, a novel neural network model is proposed to classify diabetic retinopathy. Worldwide, diabetic retinopathy has become a leading factor in acquired blindness. Convolutional neural networks (CNN) are typically used as the standard approach for identification and classification. But these networks have limits when it comes to identifying locations and global properties. ViT can overcome these limitations, but it lacks CNN's inductive bias and capacity to extract local information. As a solution, we proposed in this study that ViT and AlexNet be combined to produce a new AlexVit model. AlexNet, which produces a feature map after convolution and layer pooling. Next, the feature map is stretched into a sequence that can be input into ViT, where the location data is added and the class token is extracted as the classification result after transformer. AlexVit was trained on 3632 fundus images with different DR degrees on the Kaggle. The model's indicators significantly outperformed the traditional CNN LeNet and ViT, with kappa reaching 0.8141, 0.1423, and 0.1779 higher than LeNet and ViT. The train accuracy also significantly outperformed LeNet and ViT, reaching 0.7370, which was 0.0356 and 0.0329 higher than LeNet and ViT. All of the models and scripts referenced in this paper are available on Github.