<p>Diabetic retinopathy (DR) is a severe complication that poses a major threat to vision health and can lead to irreversible blindness without timely treatment. Therefore, accurately predicting the progression and grading of DR is crucial for enabling early intervention and formulating effective clinical treatment strategies. Traditional classification methods typically rely on single-type features, which limit their ability to capture the complex structural and pathological variations in fundus images. To overcome this limitation, we propose DSF-Net, a dual-stream fusion network integrating structural and detailed features for fundus-based Diabetic Retinopathy classification. To further enhance feature integration, we design a fusion branch composed of a cascaded Structure-Detail Fusion Block (SDFusion Block) and a Fusion Feature Extraction Block (FFE Block), which enables effective and adaptive feature fusion while mitigating redundancy and conflict. Moreover, to leverage complementary information from multiple perspectives, we introduce a Multi-View Agreement Classifier that adaptively aggregates predictions from different branches to improve classification robustness. Extensive experiments conducted on both a public dataset and our private collected dataset demonstrate that our method achieves superior performance compared to existing state-of-the-art approaches across various evaluation metrics.</p>