OptIDR combines cutting-edge deep learning with intuitive explainability to detect DR severity (0–4) from retinal fundus images, delivered via a Streamlit web demo. Early diagnosis can prevent up to 95% of vision loss—crucial for communities with limited ophthalmology access.
Diabetic retinopathy (DR) is the leading cause of adult blindness worldwide. Regular screenings are scarce in underserved areas, delaying intervention and increasing irreversible vision loss.
- Web Demo (Streamlit): Upload a fundus image ➔ get a DR stage prediction + explainable heatmap + triage recommendation.
- Models Developed:
- EfficientNetB5 (Keras): achieves 85% validation accuracy, Cohen’s QWK pre-optimization 0.9190 → post-optimization 0.9345.
- ResNeXt50 (PyTorch): reaches 83% validation accuracy and converges in 16 epochs (best val loss 0.1744).
- AI-Driven Diagnosis
- Transfer-learned CNNs (EfficientNetB5, ResNeXt50) fine-tuned on APTOS 2019.
- Explainability (Grad-CAM)
- Visual heatmaps highlight retinal lesions driving each prediction.
- Risk Triage
- Automated referral advice (e.g., "High risk: specialist within 1 week").
- Low-Bandwidth Ready
- Model quantization/pruning for efficient inference in constrained settings.
- User-Friendly UI
- Streamlit interface for instant upload → classify → visualize workflow.
| Component | Technology |
|---|---|
| Model Training | TensorFlow (Keras) / PyTorch |
| Explainability | OpenCV, Grad-CAM |
| Web Demo | Streamlit |
| Optimization | Adam |
| Model | Val Accuracy | Val QWK (pre) | Val QWK (opt) | Convergence (epochs) | Params (M) | Best Val Loss |
|---|---|---|---|---|---|---|
| EfficientNetB5 (Keras) | 85.00% | 0.9190 | 0.9345 | 35 | 30 | 0.033 |
| ResNeXt50 (PyTorch) | 83.00% | 0.8912 | 0.9215 | 16 | 25 | 0.1744 |
MIT © NEXUS