Skin Lesion Classifier
EfficientNet-B0 fine-tuned on HAM10000 — running entirely in your browser.
Loading model…
Choose or drop a dermatoscopic image
Nothing is uploaded — inference runs locally.
Nothing is uploaded — inference runs locally.
Results will appear here.
Read before trusting any output.
On a held-out, lesion-level test split this model reaches 83.1% accuracy and
0.9647 macro ROC-AUC — but it misses 36% of malignant lesions
(101 of 284). The headline accuracy is flattered by class imbalance: benign
nevi are ~67% of HAM10000. This is a research and portfolio project,
not a medical device. It is not validated for clinical use and
must never be used to make or defer a diagnosis. HAM10000 also skews toward
lighter Fitzpatrick skin types. See a dermatologist about any concerning lesion.
| Metric | Value |
|---|---|
| 7-class accuracy | 83.1% |
| Macro ROC-AUC | 0.9647 |
| Malignant sensitivity | 64.4% |
| Malignant specificity | 94.1% |
| Missed malignant lesions | 101 of 284 |