DermAssist – ResNet50 Ensemble
Model Overview
DermAssist is a dermoscopic skin lesion malignancy risk estimation model built using a 3-model ResNet50 ensemble.
The system outputs:
- Malignancy probability
- Confidence score
- Uncertainty estimation (ensemble disagreement)
This model is designed for dermoscopic image triage workflows.
Architecture
Backbone: ResNet50 (ImageNet pretrained)
Classifier Head: Linear → ReLU → Dropout (0.5) → Linear → 1 output logit
Training Details:
- Optimizer: AdamW (lr = 1e-4, weight_decay = 1e-4)
- Loss: BCEWithLogitsLoss (class-weighted)
- Early Stopping: Validation AUC
- Full fine-tuning applied
Ensemble Strategy:
Three independently trained models.
Final probability = mean prediction.
Uncertainty = standard deviation across models.
Performance
Internal Validation (HAM10000)
| Metric | Value |
|---|---|
| AUC-ROC | 0.937 |
| Accuracy | 0.86 |
| Malignant Recall | > 0.90 |
| Malignant Precision | ~0.52 |
| Benign Precision | ~0.96 |
External Validation (ISIC 2019 – Filtered Classes)
| Metric | Value |
|---|---|
| AUC-ROC | ~0.72 |
| Accuracy | ~0.57 |
Observed performance variation reflects dataset distribution differences.
Intended Use
This model is intended for:
- Dermoscopic image triage
- Educational experimentation
- Ensemble uncertainty analysis
- Explainability demonstration
This model is not intended for autonomous medical diagnosis.
Limitations
- Trained on dermoscopic images only
- Binary mapping reduces lesion taxonomy complexity
- Sensitive to dataset distribution shift
- Requires calibrated threshold tuning
Files in Repository
- resnet50_model_1.pth
- resnet50_model_2.pth
- resnet50_model_3.pth
All three models are required for ensemble inference.
Ethical Considerations
AI-based medical systems must be used responsibly.
This model does not replace dermatologist evaluation.