NailDiseaseDetection

Experiment Steps

1. Data Collection

• Collected nail images from multiple sources (public datasets + clinical samples).
• Included 22 fine-grained categories (e.g., healthy, psoriasis, fungal infection).
• Cleaned data by removing duplicates, blurry images, and mislabeled samples.

2. Data Integration

• Unified datasets into a single standardized structure.
• Created a label mapping:
  • 22 fine-grained → 10 coarse categories
• Train/val/test split: 70% / 15% / 15%
• Preprocessing:
  • Resize images → 224×224
  • Normalize pixel values
  • Augment data (random flip, rotation, brightness adjustment)

3. Model Training

• Framework: PyTorch
• CNN backbones: ResNet18, ResNet50, MobileNetV2, EfficientNet-B0
• Training setup:
  • Optimizer: Adam (lr=1e-4)
  • Loss: Cross-Entropy Loss
  • Batch size: 32
  • Epochs: 30–50
  • Early stopping to avoid overfitting

4. Model Evaluation

• Metrics: Accuracy, Precision, Recall, F1-score
• Evaluation levels:
  • Fine-grained (22 classes)
  • Coarse (10 classes)
  • Hierarchical refinement (Psoriasis vs Shape Deformities)
• Visualizations:
  • Confusion matrix
  • ROC curves

5. Explainability with Grad-CAM

• Applied Grad-CAM to highlight disease-related nail regions.
• Generated heatmaps for each prediction.
• Compared model interpretability across different CNNs.

6. Deployment Testing

• Built FastAPI backend with endpoints for prediction & Grad-CAM.
• Frontend allows users to:
  • Upload or capture an image
  • View prediction results with probabilities
  • Toggle Grad-CAM heatmap
• Tested latency and throughput under different settings.