Enhancing Diabetic Retinopathy Classification with a Hierarchical Residual Backbone and Attention Mechanism

We introduce a novel backbone architecture designed for improved diabetic retinopathy (DR) classification. Our approach focuses on two key enhancements:

1. Hierarchical Residual-like Connections: Inspired by the success of residual learning, we incorporate hierarchical residual-like connections within each single radix block of the backbone. This design significantly enhances the model's ability to capture and represent features at multiple scales, crucial for accurate DR grading.

2. Attention Force Mechanism: To further improve classification accuracy, we integrate an attention force mechanism into the model. This mechanism selectively attends to lesion-relevant features while suppressing background noise and non-lesion information. By focusing on the most informative regions of the image, the model gains a stronger ability to distinguish between different DR severity levels.

Our experimental evaluations demonstrate that the proposed model achieves superior performance compared to existing methods for DR classification. The hierarchical residual connections contribute to a richer feature representation, while the attention force mechanism minimizes the impact of irrelevant information, leading to improved classification accuracy.