AI-assisted analysis of skin lesion images
Upload a dermoscopic image of a mole or skin lesion. DermaScan estimates how likely it is to be melanoma or another skin cancer, ranks 7 possible diagnoses, and highlights the regions that drove its decision.
- Training images
- 10,015
- Diagnoses
- 7
- External test images
- 1,511
- Cancer detection AUC
- –
notebooks/train_on_colab.ipynb), then restart the app.
Analyse an image
Upload an image
A close-up dermoscopic image works best.
Drag and drop an image
or browse your files
JPEG, PNG or WebP · up to 15 MB · you can also paste an image
Your image is analysed in memory and never stored.
Or try an example
From the held-out test set; the label is the confirmed diagnosis.
Tips for a reliable result
- Use a dermoscopic image (taken with a skin microscope or dermatoscope attachment). Ordinary phone photos look very different and give unreliable results.
- Centre the lesion and fill most of the frame with it; only the central square is analysed.
- Make sure the image is sharp, evenly lit and free of hair or reflections where possible.
- One lesion per image.
Result
Upload an image to see the analysis.
Your result will appear here.
Analysing image…
Checking 4 orientations and computing the attention map.
Probability by diagnosis
Red bars mark malignant or pre-malignant categories. The malignancy probability is their sum.
From image to risk estimate in four steps
-
01
Pre-processing
The image is turned upright, cropped to the central square and resized to the network's input size, exactly as during training.
-
02
Deep neural network
A convolutional network pre-trained on millions of everyday images and fine-tuned on 7,009 dermoscopic images scores 7 diagnoses.
-
03
Robust averaging
Skin lesions have no "up", so the image is analysed in 4 flipped orientations and the predictions are averaged.
-
04
Screening decision
The probabilities of melanoma, basal cell carcinoma and actinic keratosis are added up and compared with a threshold set to catch 90% of cancers.
How well does it work?
The model was tested once, on images it never saw during training: a held-out part of HAM10000 (no lesion appears in both training and test data) and the independent official ISIC 2018 challenge test set.
Test results appear here once a trained model is loaded.
- Sensitivity
- Share of cancers that are flagged. Missing a cancer is the costliest error, so the threshold targets 90%.
- Specificity
- Share of harmless lesions correctly not flagged. Lower specificity means more false alarms.
- AUC
- How well the score separates malignant from benign lesions across all thresholds: 0.5 is guessing, 1.0 is perfect.
- Balanced accuracy
- Average correct rate over the 7 diagnoses, so rare cancers count as much as common moles (official ISIC 2018 metric).
The ABCDE rule for moles
Dermatologists use these warning signs of melanoma. If a mole shows any of them, have it checked by a doctor, whatever an app says.
Asymmetry
One half does not match the other.
Border
Edges are ragged, notched or blurred.
Colour
Several shades of brown, black, red, white or blue.
Diameter
Larger than about 6 mm (a pencil eraser), though melanomas can be smaller.
Evolving
Changes in size, shape, colour, or new itching or bleeding.
The 7 diagnoses DermaScan distinguishes
Questions and limitations
Can DermaScan tell me whether I have skin cancer?
No. It is a student research project, not a certified medical device. Only a doctor can diagnose skin cancer, usually by examining the lesion with a dermatoscope and, if needed, a biopsy. Use DermaScan to learn how AI analyses skin images, not to make health decisions.
Does it work with photos from my phone?
Not reliably. The model learned from dermoscopic images, which are magnified, evenly lit and taken in contact with the skin. Ordinary photos differ in scale, lighting and focus, so results on them are unreliable.
What is the attention map?
It is a Grad-CAM visualisation: it shows which regions most increased the score of the predicted diagnosis. It helps to spot when the model looks at the wrong thing (such as a ruler, hair or a skin marker), but it is not a medical finding.
Why is a lesion flagged even when "benign" is the most likely diagnosis?
The referral threshold is deliberately low so that about 9 in 10 cancers are flagged. The price is more false alarms: a flagged result means "worth a closer look", not "cancer".
Who was the model trained on?
HAM10000 contains images mostly from fair-skinned patients in Austria and Australia. Performance on darker skin types, on body sites that are under-represented, and on lesion types outside the 7 categories is unknown and probably worse. The model cannot say "I don't know".
What happens to my image?
It never leaves your device. The model runs inside your browser, so the image is analysed on your own phone or computer and is not uploaded anywhere. The page loads no third-party scripts or trackers.