Derma-AI: Advanced Skin Lesion Classification
An AI-powered system designed for accurate and early classification of dermatological conditions using cutting-edge deep learning fusion architecture. The classifier supports 7 skin lesion categories: Melanoma, Melanocytic Nevi, Basal Cell Carcinoma, Actinic Keratoses, Benign Keratosis-like Lesions, Dermatofibroma, and Vascular Lesions.
HAM10000 Dataset
A large dermatoscopic image dataset with 10,015 images across 7 diagnostic categories, enabling robust AI-based skin lesion classification.
Total Images
10,015
Classes
7
ViT + EfficientNet Fusion
A novel architecture that combines Vision Transformer (ViT) and EfficientNet-B0, extracting complementary features and enhancing classification accuracy.
Total Parameters
~120M
Input Size
224×224
89.9%
Accuracy
Final validation accuracy after 40 training epochs.
0.9766
F1-Score
Weighted average F1-score across all lesion classes.
0.9766
AUC-ROC
Area Under ROC Curve across all classes (One-vs-One).
0.9999
Training F1
Final training F1-score, indicating strong convergence.
✅ Trained for 40 epochs using AdamW optimizer and CosineAnnealingLR.
📉 Final Train Loss: 0.0086 | Val Loss: 0.4660
📈 Accuracy: 89.9% | F1-Score: 0.9766 | AUC: 0.9766
🔁 Used Stratified 5-Fold Cross-Validation with consistent performance across folds.
🧠 Minimal overfitting observed; model shows strong generalization.
Charts (confusion matrix, per-class accuracy, etc.) will be integrated here.
Coming soon...