Diabetic Retinopathy Detection
Grading diabetic retinopathy from retinal fundus images at clinical agreement.
A deep-learning pipeline for multi-class grading of diabetic retinopathy from retinal fundus images, benchmarking EfficientNet, ConvNeXt, and Swin-Tiny backbones and adding Grad-CAM explainability to highlight clinically relevant regions. It reaches 90.42% accuracy and a 94.17% Quadratic Weighted Kappa — indicating strong agreement with clinical grading.
Diabetic retinopathy is a leading cause of preventable blindness, but screening is bottlenecked by the shortage of ophthalmologists — and black-box models won't earn clinical trust.
Engineered pipelines across three modern backbones (EfficientNet, ConvNeXt, Swin-Tiny) for five-class DR classification, paired with Grad-CAM visualizations so predictions can be inspected against clinically relevant regions of the retina.
- 1Multi-class imbalance across the five DR severity grades.
- 2Interpretability strong enough to earn clinical trust.
- 3Benchmarking modern CNN and transformer backbones fairly.
- 90.42% accuracy with 0.94 Quadratic Weighted Kappa.
- Three-backbone benchmark: EfficientNet, ConvNeXt, Swin-Tiny.
- Grad-CAM overlays for clinical interpretability.