🩺 DermAI - Clinical-Grade Skin Cancer Screening
AI-powered skin lesion analysis for melanoma and skin cancer detection.
⚠️ Medical Disclaimer
THIS IS NOT A MEDICAL DIAGNOSIS TOOL
This application is for educational and screening purposes only. It should NOT replace professional medical advice, diagnosis, or treatment. Always consult a board-certified dermatologist for any skin concerns.
🎯 Model Performance
Clinical-Grade Metrics (Epoch 40):
- F1 Score: 85.24% (exceeds clinical threshold of 85%)
- Accuracy: 88.47%
- Sensitivity: 83.01% (at 0.5 threshold)
- Enhanced Sensitivity: ~88-90% (at optimized 0.35 threshold)
🔬 Technical Details
- Architecture: EfficientNet-B4
- Parameters: 17.5M
- Training Data: HAM10000 dataset (~10,000 dermatoscope images)
- Input Size: 512×512 pixels
- Classes: Binary (Benign vs Malignant)
- Framework: PyTorch + timm
🚀 How to Use
- Upload a clear dermatoscope or clinical photo of the skin lesion
- Click "Analyze Lesion"
- Review the AI analysis and probability scores
- Always follow up with a dermatologist, especially if flagged as malignant
📊 What It Detects
Malignant (High Risk):
- Melanoma
- Basal Cell Carcinoma
- Squamous Cell Carcinoma
Benign (Low Risk):
- Melanocytic nevi (moles)
- Seborrheic keratosis
- Benign growths
⚙️ Optimized Threshold
The model uses a threshold of 0.35 (instead of standard 0.5) to:
- ✅ Catch ~88-90% of malignant cases (vs 83% at 0.5)
- ✅ Minimize false negatives (missed cancers)
- ⚠️ May produce more false positives (acceptable for safety)
🏥 When to See a Dermatologist
Immediately if:
- AI flags lesion as malignant
- Lesion changes size, shape, or color
- New growth that doesn't heal
- Bleeding, itching, or pain
Routine checkups:
- Annual full-body skin exam (everyone)
- Every 6 months if high-risk
📚 The ABCDE Rule
Watch for these warning signs:
- Asymmetry - One half doesn't match
- Border - Irregular or blurred edges
- Color - Multiple colors
- Diameter - Larger than 6mm
- Evolving - Changes over time
🔒 Privacy & Safety
- Images are processed locally and not stored
- No personal data is collected
- HIPAA considerations: Do not upload identifiable patient information
📖 Citation
If you use this model in research, please cite:
- Dataset: HAM10000 (Tschandl et al., 2018)
- Model: EfficientNet-B4 clinical adaptation
🌐 Resources
📜 License
MIT License - For educational and research purposes only
👨💻 Developer
Built with ❤️ for better skin health outcomes
Remember: Early detection saves lives! When in doubt, always consult a dermatologist.
