MelanomaClassifier

A deep learning project for classifying skin lesions as melanoma or non-melanoma using medical images. This repository implements a pipeline that combines Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs), leveraging the PH2Dataset for training and validation.

🧠 Features

• Modular CNN and MLP models for flexible experimentation
• Custom cosine learning rate scheduler for smoother convergence
• 5-fold and 10-fold cross-validation for robust performance evaluation
• Dataset loading and preprocessing tailored to the PH2 dataset
• Support for training and inference workflows

📂 Project Structure

File/Folder	Purpose
PreProcessing/	Image and data preprocessing tools
PH2Dataset/	Contains the dataset (if not excluded)
FeatureBuilders/	Scripts to extract additional features
results/	Accuracies and recalls of all models
1D_10fold_CNN.py	10-fold CV implementation of the 1-dimension CNN
CNNClassifier.py	CNN model architecture
CNNClassifier5FoldCV.py	5-fold CV for CNN
CustomLearningRateScheduler.py	Learning rate strategy
MLPClassifier.py	MLP model
MLPClassifierGridSearch.py	MLP model with Grid Search
mod_PH2_dataset.csv	Processed metadata
read_images.py	Image reading and preprocessing

---

📦 Installation

git clone https://github.com/eSVeeF/MelanomaClassifier.git
cd MelanomaClassifier
pip install -r requirements.txt

🖼️ Dataset

The model is trained on the PH2 Dataset and organized in the following format:

MelanomaClassifier/
├── PH2Dataset/
│   └── Custom Images/
│       ├── Lesion/
│       ├── Normal/
│       └── Others/

🚀 Usage

MLP Classifier

To run the Multilayer Perceptron (MLP) classifier with grid search:

python MLPClassifierGridSearch.py

1D 10-Fold CNN

To execute the 1-dimensional CNN 10-fold:

python 1D_10fold_CNN.py

CNN Classifier

To train the Convolutional Neural Network (CNN) with 5-fold cross-validation:

python CNNClassifier5FoldCV.py

📉 Sample Results

The models achieve a competitive 82% accuracy and 91% recall on PH2 data using standard metrics. Code is modular and ready for adaptation to other medical imaging datasets.

🤝 Contributing

Contributions are welcome! Please open an issue or submit a pull request for improvements, bug fixes, or feature requests.

License

This project is licensed under the MIT License, feel free to use and modify for non-commercial purposes.