AI-powered skin awareness on your phone. Photograph a spot. Learn about your skin. Your photos never leave your device.
Research validation metrics (Source: scripts/combined-training-results.json). Not clinical diagnostic performance claims.
Americans will develop skin cancer by age 70. Early detection is the single biggest factor in survival.
5-year survival rate for melanoma caught early. Drops to 35% when found late. The difference is time.
Mela is open-source, runs entirely on your phone, and costs nothing. Because skin awareness shouldn't be a luxury.
Mela is a Progressive Web App. Visit the link on your phone and add it to your home screen. Works offline.
Visit mela-app.vercel.app on Safari or Chrome. Tap "Add to Home Screen" for the full app experience.
Good lighting. Six inches away. Center one spot in the frame. The app checks image quality automatically.
The analysis shows pattern similarity to reference categories. This is educational, not diagnostic. Always consult a dermatologist for medical concerns.
The AI model runs entirely on your device. Your photos are never uploaded to any server. There are no accounts, no tracking, no data collection. We can't see your images because they never leave your phone.
A single neural network isn't trustworthy enough for skin analysis. Mela uses a 4-layer ensemble with safety gates that catch what any single model misses.
Custom Vision Transformer trained on 37,484 dermoscopic images from HAM10000 and ISIC 2019. 95.97% melanoma sensitivity on external validation.
20-feature logistic regression with every weight cited to published dermatology literature. Cross-checks the AI's pixel-level assessment with clinical knowledge.
ABCDE criteria, Total Dermoscopy Score, 7-point checklist. The same tools dermatologists use in-office, computed from computer vision features.
Age, sex, body location, skin type. Bayesian adjustment of risk using epidemiological prevalence data. A mole on a 60-year-old's back means something different than on a 20-year-old's arm.
Hard-wired overrides that cannot be disabled: if 2+ ABCDE criteria are suspicious, melanoma probability floors at 15%. If Total Dermoscopy Score exceeds 5.45, malignant probability floors at 30%. These gates run after the ensemble and override lower estimates.
Previous versions had fabricated claims. An internal audit caught them. Now every performance number must trace back to a JSON evidence file. No exceptions.
These metrics reflect research validation on reference datasets. They do not represent clinical diagnostic performance.
We publish our failures more prominently than our successes. You deserve to know what this tool cannot do.
Mela is a skin awareness tool. It does not diagnose, screen for, or detect any disease. It does not replace a dermatologist's evaluation.
Performance on phone camera photos has not been clinically validated yet.
Melanoma sensitivity drops significantly for darker skin tones (IV-VI). We're working on it.
~20% of benign moles get flagged. The model errs on the side of caution to encourage professional consultation.
Mela's source code, training data references, evidence files, and architecture decisions are all public on GitHub. We believe health-related AI should be auditable.
Browse the full source, read the architecture decisions (13 ADRs), review the evidence files, and see exactly how every classification decision is made.
github.com/stuinfla/MelaWe need help from dermatologists, ML engineers, designers, and anyone who cares about making skin awareness accessible. Check the README for contribution guidelines.
How to contributeEnter your phone number and we'll text you the Mela app link. Open it on your phone, add to home screen, done.
This opens your phone's text messaging app with a pre-filled message. We don't store your number.
This is an open-source project built for you. Every piece of feedback helps us improve accuracy, add features, and catch issues we might miss.