Melanoma Detection

Disclamer: For educational purposes only

Melanoma CancerDetection Algorithm

Upload an image of a lesion on your skin and get an estimate of the probability of melanoma cancer. This algorithm was made as a student project on the IT-University in Copenhagen.

Melanoma cancer cell

PAD-UFES-20 Dataset

The algorithm is trained on the comprehensive PAD-UFES-20 dataset, featuring thousands of high-resolution dermatological images.

2298 imagesBiopsy-proven labels26 features
Womens shoulder with freckles on it

ABC model

Melanoma lesions are marked by asymmetry, irregular borders, and color variation—known as the ABC model. We tested methods to remove hair and extract these features with computer vision algorithms.

MaskingHair removalABC features
Bar chart with increasing white bars

Training and testing

Logistic regression delivered the best AUC and F1 score among tested models. To handle class imbalance, we used oversampling and stratified k-fold cross-validation. Read the research report for details.

AUC 95% CI: 0.579 – 0.941Oversampling

Try for ourself

Upload an image of a lesion on your skin. Within a few seconds you will get back estimates of the probability of it being melanoma cancer.