DermaStratif - Multiclass Lesion Stratification and Diagnosis

This repository implements EfficientNet-B0 for skin lesion classification using three different fine-tuning methods:

1. Base Model Training
2. Adapter Fine-Tuning
3. LoRA Fine-Tuning

The objective is to classify skin lesion images into 8 disease categories while improving accuracy and efficiency using advanced fine-tuning techniques.

---

Project Overview

Base Model

• Model: EfficientNet-B0 pre-trained on ImageNet.
• Classifier: Fully connected layers added to support 8 disease categories.
• Validation Accuracy: 74%
• Training Time: ~4 hours on a MacBook Pro M1 GPU.
• Description: The base model was trained with all layers unfrozen to fine-tune for skin lesion classification. It serves as the foundation for further improvements using Adapter and LoRA fine-tuning.

Adapter Fine-Tuning

• What Are Adapters?
  • Small, lightweight layers added to the model's classifier for task-specific fine-tuning.
  • Parameters outside these adapters remain frozen, reducing computational cost.
• Validation Accuracy: 76%
• Training Time: ~1 hour.
• Description: Adapter fine-tuning improves resource efficiency while achieving better accuracy than the base model.

LoRA Fine-Tuning

• What Is LoRA?
  • Low-Rank Adaptation (LoRA) introduces task-specific low-rank matrices into the model's classifier.
  • Significantly reduces memory usage while maintaining high accuracy.
• Validation Accuracy: 80%
• Training Time: ~45 minutes.
• LoRA Configuration:
  • Rank (r): 8
  • Scaling Factor (alpha): 32
• Description: LoRA fine-tuning yielded the highest validation accuracy with minimal computational cost.

---

Disease Classes

The model classifies skin lesions into the following 8 categories:

1. Melanoma: High-risk skin cancer with irregular moles.
2. Basal Cell Carcinoma: Pearly or waxy bumps on the skin.
3. Squamous Cell Carcinoma: Firm, red nodules with scaly, crusted surfaces.
4. Actinic Keratosis: Rough, scaly patches commonly found on sun-exposed skin.
5. Benign Keratosis: Non-cancerous skin growths appearing as brown or black patches.
6. Dermatofibroma: Small, firm nodules caused by minor injuries like insect bites.
7. Vascular Lesion: Marks or growths caused by abnormal blood vessels.
8. Melanocytic Nevus: Common moles that are small and pigmented.

---

Application Details

How to Use

1. Download the Repository: Clone or download the repository files:

   git clone <repository-url>

2. Prepare Required Files:

• Download the application folder and the saved model folder.
• Update the saved_model_path in app.py to the path of your downloaded model.

3. Install Dependencies: Install the required Python libraries:

• pip install -r requirements.txt

4. Run the Application: Start the Flask server by running:

• python app.py

5. Access the Application: Open your browser and navigate to:

• http://127.0.0.1:5000/

6. Upload an Image: Use the upload form to submit a skin lesion image and view the predictions.

Dependencies

Install all dependencies using the following command:

• pip install -r requirements.txt

Key Libraries:

• Flask
• PyTorch
• torchvision
• Pillow
• scikit-learn

Future Improvements

• Support for larger models (e.g., EfficientNet-B3 or ViT) for higher accuracy.
• Provide a live API for disease prediction.
• Enhance the UI with interactive features and visualization tools.