🧬 Skin Disease CNN Classification with Explainable AI (Grad-CAM)

An end-to-end deep learning project for automated skin disease classification using Convolutional Neural Networks, Transfer Learning, and Explainable AI (XAI) techniques. This repository covers the entire ML lifecycle — from exploratory data analysis (EDA) and dataset inspection to model training, fine-tuning, evaluation, explainability with Grad-CAM, and deployment using Streamlit.

Skin diseases often present visually similar symptoms, making accurate diagnosis challenging even for trained professionals. This project aims to assist dermatological screening by leveraging deep learning models capable of learning discriminative visual features from skin images while also providing visual explanations for their predictions.

• Classification of 8 skin disease categories
• Transfer Learning using ResNet18
• Class imbalance handling with weighted loss
• Accuracy improved from 67.09% → 88.03% → 95.73%
• Grad-CAM for model interpretability
• Deployed as an interactive **Streamlit web ap

⚠️ Disclaimer: This system is intended for educational and research purposes only and should not replace professional medical diagnosis.

This project uses the Skin Disease Dataset available on Kaggle:

https://www.kaggle.com/datasets/subirbiswas19/skin-disease-dataset

The dataset is organized into two primary directories:

Dataset/
├── train_set/
│   ├── BA-cellulitis/
│   ├── BA-impetigo/
│   ├── FU-athlete-foot/
│   ├── FU-nail-fungus/
│   ├── FU-ringworm/
│   ├── PA-cutaneous-larva-migrans/
│   ├── VI-chickenpox/
│   └── VI-shingles/
└── test_set/
    └── (same 8 classes)

Both training and testing sets consistently contain 8 skin disease classes, ensuring a clean and reliable evaluation pipeline.

1. BA-cellulitis
2. BA-impetigo
3. FU-athlete-foot
4. FU-nail-fungus
5. FU-ringworm
6. PA-cutaneous-larva-migrans
7. VI-chickenpox
8. VI-shingles

• The dataset is relatively balanced, though some mild imbalance exists (e.g., BA-impetigo).
• Sample images show high variability in lighting, background, and image quality, reflecting real-world medical data.
• No missing or corrupted image files were detected.

• Custom CNN trained from scratch
• Achieved 67.09% accuracy
• Served as a reference baseline

• Backbone: ResNet18 (pre-trained on ImageNet)
• Classifier Head: Custom fully connected layers
• Loss Function: CrossEntropyLoss with class weights
• Optimizer: Adam

To mitigate bias toward majority classes:

• Class weights were computed (range: ~0.85 – 1.45)
• Integrated directly into the loss function
• Resulted in significantly improved macro-averaged metrics

• Only classifier head trained
• Epochs: 15

Performance:

• Accuracy: 88.03%
• Macro Recall: 88.03%
• Macro F1-score: 86.81%

➡️ Large improvement over baseline CNN

• Unfroze layer4 of ResNet18
• Applied learning rate scheduler
• Continued training for 15 epochs

➡️ Significant improvement over both baseline CNN (67.09%) and initial transfer learning (88.03%).

• 100% Recall: BA-impetigo, FU-nail-fungus, VI-chickenpox

• High F1-scores:

  • FU-nail-fungus: 98.46%
  • VI-chickenpox: 98.51%

• Excellent generalization across most classes

• FU-ringworm shows slightly lower recall (~83%)
• BA-impetigo precision (~87%) indicates some false positives

• Confusion matrix revealed minor confusion between visually similar fungal infections
• Underperforming classes benefit most from additional data diversity

The trained model is deployed using Streamlit:

• Upload a skin image
• Get predicted disease, confidence score, and medical suggestion
• Optimized for Python 3.10 + TensorFlow 2.13+ compatibility

• Training vs Testing loss curves
• Accuracy, Recall, F1-score plots
• Confusion matrix
• Sample predictions

In medical AI, interpretability is critical. Grad-CAM (Gradient-weighted Class Activation Mapping) provides visual explanations by highlighting image regions that most influenced the model’s prediction.

• Target Layer: ResNet18 → layer4[1].conv2
• Gradients + feature maps used to compute class-discriminative heatmaps
• Heatmaps overlaid on original images

• Heatmaps focus on lesion-affected regions
• Minimal attention to background or artifacts
• Confirms learning of clinically meaningful features

• Actual: VI-shingles
• Predicted: BA-impetigo

Grad-CAM showed:

• Strong focus on lesion region
• Misclassification due to visual similarity, not random behavior

➡️ Indicates correct localization but imperfect feature interpretation.

• Grad-CAM verifies the model relies on pathological regions
• Helps identify reasons for misclassification
• Useful for detecting dataset bias or shortcut learning

Current observations suggest low bias risk, but broader Grad-CAM analysis is recommended.

🧠 Explainability is not optional in medical AI — Grad-CAM bridges the gap between model performance and human trust.

• Python
• PyTorch
• Torchvision (ResNet18)
• NumPy, Matplotlib, Seaborn
• Streamlit

Abhay Singh BCA (AI & ML + Data Science) Vivekananda Institute of Professional Studies

🔗 GitHub: https://github.com/AbhaySingh71 🔗 LinkedIn: https://www.linkedin.com/in/abhay-singh-050a5b293/

⭐ If you find this project helpful, consider giving it a star!