Diabetic Retinopathy Screening Assistant: An AI-Powered Retinal Fundus Analysis System Utilizing a CNN-Transformer Hybrid Architecture

Diabetic retinopathy (DR) remains a leading cause of preventable vision loss globally. While early intervention is crucial for preserving sight, a severe shortage of available ophthalmologists—especially in resource-constrained regions—creates a massive bottleneck for systematic screening. To address this clinical gap, we introduce DR Detect AI, an end-to-end, web-based screening assistant. At the core of this system lies a novel hybrid deep learning architecture that fuses the localized feature extraction capabilities of a Convolutional Neural Network (EfficientNet-B3) with the global spatial awareness of a Vision Transformer encoder. Furthermore, unlike traditional multi-class classification models that treat all diagnostic errors equally, our approach integrates the Consistent Rank Logits (CORAL) framework to perform ordinal regression. This specific design choice ensures the model mathematically respects the natural severity progression of DR (from mild to proliferative), which substantially improves diagnostic consistency at borderline stages. The trained model is deployed via a high-performance FastAPI backend and connected to a React-based clinical dashboard. This interface equips medical professionals with real-time severity grading, interactive disease progression forecasting, and automated clinical reporting. Early evaluations indicate that our hybrid ordinal model achieves competitive accuracy on standard benchmark datasets, resulting in a highly scalable, accessible, and interpretable tool for real-world ophthalmic triage.