Abstract
Diabetic retinopathy represents a primary cause of preventable blindness worldwide, driving an urgent demand for accurate and scalable AI-based screening solutions. This paper comprehensively reviews recent deep learning developments, examining progress in CNNs, transformers, data balancing techniques, and interpretability frameworks while acknowledging persistent challenges in cross-population generalizability and practical clinical deployment. We introduce TriFusionNet, an innovative architecture that fuses ResNet-50, DenseNet-121, and ConvNeXt-Tiny through attention-guided integration and ordinal regression methodology. Kaggle APTOS 2019 evaluation demonstrates 98.71% accuracy, surpassing established state-of-the-art baselines, with Grad-CAM visualizations confirming lesion-focused predictions for enhanced clinical transparency and physician confidence.
