Retina AI Heart Risk Predictor

I need a working prototype that takes retina-fundus images, runs them through a deep-learning pipeline, and returns a clear heart-attack risk score—low, medium, or high—together with a confidence percentage. The core of the project is a convolutional model (CNN, RetinaNet, or whichever architecture proves most accurate after experimentation) trained on a suitably annotated dataset.

The finished solution must run end-to-end inside a web-based interface: the user drags a scan into the browser, the image is processed server-side (Python, TensorFlow or PyTorch, OpenCV as needed), and the prediction appears instantly on screen accompanied by a heat-map or attention overlay that highlights the retinal regions driving the decision.

For hand-off, please include:
• Clean, well-commented source code and model weights
• The web UI (HTML/CSS/JS or a lightweight framework such as Streamlit/FastAPI + React) ready to deploy on a standard cloud VM
• A short README covering environment setup, dataset preparation, and instructions for retraining or fine-tuning

Accuracy benchmarks aren’t fixed yet, but the model should outperform naive baselines and show sensible ROC/AUC on a held-out test set. I’ll supply or help locate retina datasets; advise if additional labeling is required. Continuous collaboration is expected until the tool is reproducible on my machine and running smoothly online.