Multi-Attention Residual Network (MARN) for Interpretable 5-Class and Referable Diabetic Retinopathy Detection on Noisy EyePACS Data

Diabetic retinopathy (DR) is a leading cause of preventable blindness, and deep learning systems have achieved high performance on large, curated fundus datasets such as EyePACS. However, most studies focus either on global accuracy or on referable DR detection, with limited attention to model calibration, interpretability, and robustness on noisy public data. This work proposes a Multi-Attention Residual Network (MARN) built on EfficientNet-B0 for joint DR grading and referable DR screening on the resized EyePACS Kaggle dataset. The backbone is followed by a residual fully connected head with dropout, trained using class-balanced sampling and cross-entropy loss. Performance is evaluated on a $\mathbf{5}$-class grading task and a clinically relevant binary task (referable $\text{DR} \geq$ moderate vs non-referable). On a 6081 -image subset, the proposed MARN improves strict 5-class validation accuracy from 0.4618 to 0.4881 and macro-F1 from 0.4905 to 0.5195 over a strong EfficientNet-B0 baseline. For referable DR, MARN attains 0.780 accuracy with sensitivity 0.776 and specificity 0.783, slightly increasing specificity compared to the baseline while maintaining high sensitivity. Class-wise analysis shows the largest $\text{F1}$ gains for Severe and Proliferative DR, where MARN achieves ROC AUCs of 0.865 and 0.915, respectively. Grad-CAM visualizations confirm that predictions are driven by lesion-rich regions, and t-SNE embeddings indicate tighter clustering of late-stage DR features under MARN. These results demonstrate that the proposed architecture provides modest but consistent improvements in grading performance, strong detection of sight-threatening DR, and clinically meaningful explanations on a noisy public benchmark, supporting its use as an interpretable decision-support tool rather than a purely black-box classifier