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

We propose a novel backbone for the classification of diabetes retinopathy that enhances the multi-scale representation ability of the model. This is achieved by constructing hierarchical residual-like connections within each single radix block. Our model also includes an attention force mechanism that suppresses non-lesion feature information and enhances the learning of typical lesion features, thereby improving the learning of low-level feature information. As a result, the classification performance of the model is further improved. The experimental results demonstrate that our proposed method significantly enhances the accuracy of DR model classification.