Skin Cancer Detection System

A deep learning-based system for detecting and classifying skin cancer using the HAM10000 dataset. This project implements a CNN architecture with advanced techniques to achieve high accuracy in skin lesion classification.

Project Overview

This system uses deep learning to classify skin lesions into different categories of skin cancer, providing a reliable tool for early detection and diagnosis assistance.

Key Features

• Multi-class classification of skin lesions
• Advanced data augmentation techniques
• Transfer learning with pre-trained models
• Ensemble learning approach
• Real-time prediction capabilities
• Comprehensive model evaluation metrics

Technologies Used

Core Technologies

• Python 3.8+
• TensorFlow 2.x
• Keras
• OpenCV
• NumPy
• Pandas
• Scikit-learn

Deep Learning Models

• EfficientNetB0 (Base model)
• ResNet50 (Ensemble model)
• Custom CNN Architecture

Data Processing

• Albumentations for advanced image augmentation
• PIL for image processing
• Matplotlib for visualization

Project Structure

skin_cancer_detection/
├── src/
│   ├── data_loader.py      # Data loading and preprocessing
│   ├── model.py           # Model architecture definitions
│   ├── prepare_data.py    # Data preparation and augmentation
│   ├── train.py           # Training pipeline
│   └── utils.py           # Utility functions
├── notebooks/
│   └── skin_cancer_analysis.ipynb  # Analysis and visualization
├── data/                  # Dataset directory (not tracked in git)
├── models/               # Saved model checkpoints
├── requirements.txt      # Project dependencies
└── README.md            # Project documentation

Methodology

1. Data Preprocessing

• Image resizing and normalization
• Advanced data augmentation:
  • Random rotations and flips
  • Color jittering
  • Random brightness/contrast
  • Elastic transformations
  • Cutout augmentation

2. Model Architecture

• Base model: EfficientNetB0 with custom head
• Ensemble model: ResNet50
• Custom CNN with:
  • Batch normalization
  • Dropout layers
  • Global average pooling
  • Dense layers with ReLU activation

3. Training Strategy

• Transfer learning with fine-tuning
• Learning rate scheduling
• Early stopping
• Model checkpointing
• Cross-validation

4. Performance Optimization

• Mixed precision training
• Gradient clipping
• Learning rate warmup
• Model ensembling

Model Performance

The system achieves:

• Training accuracy: ~95%
• Validation accuracy: ~92%
• Test accuracy: ~90%
• F1-score: ~0.89
• Precision: ~0.91
• Recall: ~0.88

Setup and Installation

1. Clone the repository:

git clone https://github.com/kailhashed/Skincancer_detection.git
cd Skincancer_detection

2. Create and activate virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Download the HAM10000 dataset and place it in the data/ directory

Usage

1. Prepare the data:

python src/prepare_data.py

2. Train the model:

python src/train.py

3. Use the trained model for predictions:

from src.model import load_model
from src.utils import preprocess_image

model = load_model('models/best_model.h5')
prediction = model.predict(preprocess_image('path_to_image.jpg'))

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• HAM10000 dataset creators and contributors
• TensorFlow and Keras teams
• Open source community for various tools and libraries