About DermaMarker
DermaMarker uses deep learning to classify skin lesions across 7 clinical categories, combining a fine-tuned neural network with AI analysis to provide educational information about skin conditions.
How it works
- 1. Upload — You provide a clear photo. It's resized to 224×224 and normalized using ImageNet statistics.
- 2. Classify — An EfficientNet-B0 model fine-tuned on the HAM10000 dataset produces probability scores for each category.
- 3. Report — An agent generates bilingual summaries, clinical notes, and recommendations from the classification and triage results.
The HAM10000 dataset
HAM10000 (“Human Against Machine with 10,000 training images”) is a multi-source dermatoscopic image collection from the Medical University of Vienna and Harvard. It contains 10,015 images across 7 categories, collected over 20+ years from two clinical sites. More than 50% of diagnoses are confirmed by histopathology.
Melanocytic nevi — 6,705 (67%)
Melanoma — 1,113 (11%)
Benign keratosis — 1,099 (11%)
Basal cell carcinoma — 514 (5%)
Actinic keratoses — 327 (3%)
Vascular lesions — 142 (1.4%)
Dermatofibroma — 115 (1.1%)
Tech stack
- EfficientNet-B0 — transfer learning, fine-tuned on HAM10000
- Rule-based agent — bilingual summaries and triage (no paid API)
- Next.js + Vercel — frontend and deployment
- FastAPI + Railway — Python ML inference backend
- Supabase — Postgres database and object storage
Disclaimer
DermaMarker is an educational tool and research prototype. It is not a certified medical device and should never replace professional medical advice, diagnosis, or treatment. If you notice changes in a mole or skin lesion, consult a healthcare professional.