Enhanced Diabetic Retinal Classification with Improved ResNest Backbone

In recent years, deep learning-based diabetic retinal classification has shown remarkable results. Several studies have improved existing models, such as Qiong Li et al. [6] who enhanced the AlexNet model with another batch normalization layer and migration learning strategy, achieving 93% accuracy. Kermany[7] used the migration learning algorithm to train the InceptionV3 network on the ImageNet dataset, achieving high diagnostic accuracy and solving the problem of small data sets. Li et al. [8] developed a deep learning algorithm for DR detection based on color fundus images with high accuracy. De Fauw et al. [9] proposed a two-stage deep neural network model for OCT image diagnosis, achieving an accuracy of 95%. Hassan et al. [10] proposed a deep convolutional neural network and a structural tensor based segmentation framework (CNN-STSF) for automatically segmenting eight retinal layers. However, some methods have shortcomings, such as using overly complex backbones that may result in low inference time or ignoring the semantic information contained in deeper feature maps, leading to a deterioration of robustness [11]. To address these issues, we improved the existing ResNest backbone by constructing hierarchical residual-like connections within each single radix block, enhancing the multi-scale representation ability with fewer parameters and calculation costs.