Model notes

A small research demo for explainable DR screening.

The interface keeps prediction, confidence, and heatmap together so each result reads as one report.

Dataset

Trained on the APTOS 2019 Blindness Detection dataset, with 3,662 retinal fundus images labelled 0-4 by clinicians for DR severity.

Model

EfficientNet-B3 at 300px, ImageNet-pretrained and fine-tuned for 5-class classification with Ben Graham preprocessing, softened class weights, and EMA weights.

Explanation

Grad-CAM is generated over the final convolutional layer so each report can show the image regions that shaped the selected class.

Limits

Performance may not transfer across cameras, populations, or image quality. Retra does not detect other ocular disease, and confidence scores are not calibrated disease probabilities.

Use

Retra is a research and educational demo. It is not a medical device and must not be used for diagnosis or treatment decisions.