Run the model on your own images
Try your own image
Choose an image, see what the model says about it, then see how familiar it is at each stage of the model.
Research demo only. Not medical advice. The model learned from photos taken through a skin magnifier (a dermatoscope). A phone photo of your arm is not like those photos, so the model may give a high output probability for something it has never really seen. Think of someone who has only ever driven in one town: they can still give directions somewhere new, sound sure, and be wrong. If a lesion on your skin worries you, see a doctor.
Your image
Stays on your computerDrop an image here, or choose a file.
Real images from a public skin-lesion collection (ISIC) that the model did not learn from. Chosen to show interesting cases, so they are not typical.
The first run downloads the model and its library of training images to your browser. Your image is not sent anywhere.
Original model output probability vs our derived familiarity index
Not medical adviceChoose an image to see the results
The model’s answer, its output probability and the familiarity scores will appear here.
Model’s answer
Correct answer (from the dataset)
How familiar is your image to the model?
Early stages
Output probability
Familiarity bands are not available for this network yet. Its raw scores are under Raw scores and stage details.
Nearest training images
Each dot is a training image, flattened onto a two-dimensional map. Your image is the large bullseye. Dots close to it look alike to the model, but flattening distorts distances. The exact distances are listed under More map settings.
All training images
Click a dot to see it, or step through the nearest images below.
How to read the map. Each dot is one training image, placed so that images the model treats as alike sit close together; the map is a flattened picture, so distance on screen is only a rough guide. Dots are coloured as Benign answers or Malignant answers by the label the model gave each image, and diamond outlines mark Past mistakes, the training images it got wrong. Gold rings show the Nearest training images to Your image, found with the full set of numbers the model uses rather than with the flat map. The familiarity score is not the chance the answer is right.
More map settings
Each density = 1 / (kth distance + 0.001). C = W / (W + B + E + ε); G is shown separately.
Raw scores and stage details
The probability the network gives its top answer.
Measured at the model’s last stage.
Familiarity (from 0 to 1). Your image is compared with the images the model learned from. A score near 1 means your image sits mostly among training images the model labelled the same way and usually got right. A score near 0 means it sits nearer training images with the other label, or images the model got wrong. It is not the chance that the answer is right.
A high familiarity score does not mean the answer is right. It only means the image looks like images the model has seen before.
Familiarity scores at each stage
Each row shows the same image at a different stage inside the model, from early to late. The raw scores (W, B, G, E) use a different scale at each stage, so compare them within a row, not between rows.
| Stage | W | B | G | E | C |
|---|
This model is a recovered copy and is still being checked: we have not yet confirmed that its saved library of training images matches it exactly. About this version
What am I looking at?
The model gives one of two answers: benign (harmless) or malignant (cancerous). The familiarity score asks a different question: does your image look like the images the model learned from, and how did the model label those images?
Read how the two results differ →