Multi-Stage Grading of Diabetic Retinopathy Severity Using EfficientNet-B0 Architecture on Retinal Fundus Images

Abstract

Diabetic eye damage stays the top reason people lose sight needlessly around the world, especially those living with diabetes for years. Spotting it early helps stop harm from getting worse - however, checking eye pictures by hand takes lots of effort, can be shaky due to human error, and needs trained eye doctors. In recent times, using smart algorithms inspired by brain networks has boosted how well computers detect this condition - and let systems work on bigger scales. Here’s a new algorithm built on a shifting design, based on EfficientNetB0, sorting diabetic eye disease into five levels (from 0 to 4). A large set of eye images - more than 35,108 sharp ones from Kaggle - was used for training. To balance categories while improving adaptability, extra image tweaks were added during learning; meanwhile, Adam helped speed up progress. Our model hit an overall score of 82.34%, matching results seen in similar multi-class diabetic retinopathy studies. Results show small CNN setups can still deliver solid performance for medical-grade spotting. This pushes forward low-cost, reliable, and expandable automatic screening systems suited for real clinics.