Enhanced Diabetic Retinopathy Assessment using MIL-ViT: A Synergistic Approach for Accurate Grading and Clinical Adoption

The worldwide incidence of diabetic retinopathy (DR) remains one of the main reasons for blindness, so effective preventive measures and early interventions become vital. The current grading systems depend heavily on ophthalmologists for evaluation while maintaining qualitative standards with potential observer disagreement. This research design builds an enhanced network framework that combines MIL and ViT to solve existing challenges and achieve higher grading accuracy. The MIL-ViT model, with its enhanced features, now uses ViT attention systems to find distant patterns between ophthalmological images both at patch and image scales. The effectiveness of the EfficientNetB3 architecture exceeded other models tested during augmentation due to its exceptional 93.09% accuracy. The integration of Streamlit into the model framework produces an interface that enables medical staff to submit retinal images for automatic DR grading. This user-friendly interface ensures real-time accessibility and simplifies the diagnostic workflow, enhancing the efficiency of diabetic retinopathy screening. The proposed framework offers a scalable solution for large-scale screening, improving early detection rates.