Computer vision
Retra
A diabetic retinopathy screening platform that pairs severity predictions with readable Grad-CAM reports.
The project
Retra reads retinal fundus images, predicts diabetic retinopathy severity, and makes the reasoning visible through confidence outputs and Grad-CAM heatmaps. The whole experience is built around turning a computer vision result into a report someone can actually inspect. It grew out of a Samsung Solve for Tomorrow proposal.
01
Explainable predictions
Shows severity, confidence, and Grad-CAM overlays together instead of handing back an opaque classification.
02
Image pipeline
Uses Ben Graham preprocessing and an EfficientNet-B3 classifier tuned for retinal fundus imagery.
03
Readable reporting
Packages the model output into a screening report built for faster review and comparison.
Built with
Primary
- Python
- PyTorch
- Hugging Face
- OpenCV
- Next.js
- FastAPI
- TypeScript
Other
- EfficientNet-B3
- timm
- Grad-CAM