Mela begins with the world's best publicly available dermoscopy datasets — HAM10000 from the Medical University of Vienna and ISIC 2019 from institutions worldwide. 37,484 images of real skin lesions, each diagnosed by a dermatologist. Melanoma, basal cell carcinoma, benign moles, vascular lesions — the full spectrum of what your skin can show.
0Training Images
0Lesion Types
0HAM10000
0ISIC 2019
Every image is a real clinical case. Every classification was confirmed by a specialist.
Teaching an AI to See What Doctors See
We trained a Vision Transformer — the same architecture behind the world's best image recognition — with a critical innovation: focal loss. Standard AI treats all mistakes equally. Ours treats missing a melanoma as 6x worse than misclassifying a mole. Because missing cancer is the one mistake that matters most.
85.8MParameters
6.0xMelanoma Weight
3.3 hrsTraining Time
5Epochs
Trained on an Apple M3 Max. No cloud GPU farm. No million-dollar compute budget. Just the right architecture, the right loss function, and the right data.
Focal loss forces the model to focus on the hardest, most dangerous cases first.
Tested, Refined, and Validated
We didn't stop at one model. We continuously tested against multiple datasets and independent models, refining the architecture to maximize the information that matters: catching real cancers while minimizing false alarms that cause unnecessary worry. The final model was validated on 3,901 images it had never seen — from different cameras, hospitals, and countries.
95.97%Melanoma Sensitivity
0.960AUROC
98.3%All-Cancer Sensitivity
3,901External Test Images
Every number is measured, not claimed. Every metric cites its evidence file. An internal audit ensures nothing is overstated.
Cross-dataset validated: tested on data from cameras and institutions the model never trained on.
327MB to 85MB. Fits in Your Pocket.
A full AI model is too large for a phone. We used INT8 quantization via RuVector Format (RVF) to compress it to 85MB — small enough to download once and cache permanently. After the first visit, Mela loads instantly from your phone's local storage. No waiting. No re-downloading. Ever.
Raw Model
327 MB
RVF INT8
85 MB
After Cache
0 MB transfer
327MBOriginal
85MBCompressed
InstantAfter First Load
RuVector Format: self-contained, portable, runs entirely via WebAssembly in your browser.
Your Phone. Your AI. Completely Private.
The entire analysis engine runs inside your browser via WebAssembly. Your skin photo never leaves your device — not to a server, not to a cloud, not to anyone. This isn't a privacy policy. It's how the app is built: the data physically cannot leave your phone.
Mela
95.97% sensitivity
100% on your phone
0Photos Sent Anywhere
0Cloud Dependencies
100%Works Offline
InstantAfter First Visit
No internet needed after the first download. Analyze your skin anytime, anywhere.
What Happens When You Scan
When you photograph a skin spot, Mela runs a research-grade analysis pipeline entirely on your phone. It segments the area, measures visual features, runs the AI model, checks safety gates, and translates everything into educational information you can learn from.
Your Photo→Quality Check→Find the Lesion→Measure Features→AI Classification→Safety Check→Risk Score→Your Result
Mela doesn't just give you a label. It tells you what it found, why it matters, and what to do next — in plain language anyone can understand.
3AI Layers
2Safety Gates
5Risk Levels
PlainEnglish Results
Every step runs on your phone. Nothing is sent to any server. Ever.
Because Catching It Early Changes Everything
Melanoma caught early has a 99% five-year survival rate. Caught late, that drops to 30%. The difference is awareness. Today, professional analysis tools cost $7,000 and require proprietary hardware. That means skin awareness is a luxury available to well-funded clinics in wealthy countries. Mela is designed to change that — as an educational tool that helps you understand your skin.
95.97%Sensitivity
$0Cost
100%Private
OpenSource
Mela puts AI-powered skin awareness in the hands of anyone with a smartphone. Open source. Free forever. Educational tool, not a medical device.