Enhanced Backbone for Diabetes Retinopathy Classification: Multi-Scale Representation and Attention Mechanism

Our proposal presents a backbone for classifying diabetes retinopathy, which includes hierarchical residual-like connections within each single radix block to enhance the model's multi-scale representation ability. Additionally, our model incorporates an attention force mechanism to suppress non-lesion feature information and improve the learning of typical lesion features, ultimately improving the model's classification performance. Through experimentation, we have found that our proposed method significantly improves the accuracy of DR model classification.