Live demo

Run the model on a real image.

This page talks to a Python server on your own machine. It loads the model the six hospitals trained together, runs one dermoscopy image through it and shows what it predicts across all eight classes — including when it gets the answer wrong.

Connecting to the model…

What you are looking at

The same model the results pages report.

Nothing here is a mock-up: the predictions come from the FedAvg run on the natural split, loaded from a saved checkpoint.

Six hospitals, one model

The classifier head was trained by FedAvg across the six real Fed-ISIC2019 centres. No hospital sent an image anywhere; each sent only its updated weights.

Frozen features, trained head

An ImageNet DenseNet-121 turns each image into 1,024 numbers. Only the final layer is federated — which is what makes the whole study runnable on a laptop.

Held-out images

The eight images on this page come from the test split, chosen by class and hospital before the model was run. Two are Dermatofibroma and two are Vascular lesion: the rare classes this project is about.

Read it honestly

This is a research prototype on a public dataset, not a medical device, and it must not be used on anyone's skin. The model is right about two times in five across the eight classes — better than chance, far below clinical use — so the true label is always shown beside the prediction rather than only when it agrees. Uploaded photos are further from the dataset than the samples are: ordinary phone pictures are not dermoscopy images, and the model will answer confidently anyway.

See how it was measured The whole system, end to end