Skin Cancer Classification - Fusion_Model

This repository contains the official TensorFlow implementation for the paper: "Explainable Depth-Wise and Channel-Wise Fusion Models for Multi-Class Skin Lesion Classification"

Our work presents a systematic investigation into deep feature fusion models for multi-class skin lesion classification. This repository provides the code to train our proposed architecture, FM6, evaluate their performance, and generate the Explainable AI (XAI) visualizations (Grad-CAM and SHAP) as presented in the article.

---

Repository Structure

• FusionModel.py: The main script for loading data (HAM10000), defining the model architectures, training, evaluation, and generating XAI plots.
• GradCAM_module.py: A utility module containing the class and functions for generating Grad-CAM heatmaps.
• CM_module.py: A utility module for plotting the confusion matrix (Note: you will need to add your plot_confusion_matrix function to this file).
• requirements.txt: A list of all Python dependencies required to run the code.
• LICENSE: The MIT License file.

---

Setup and Installation

1. Clone the Repository

git clone https://github.com/Radhwan-Saleh/SkinCancerFusionModel.git
cd SkinCancerFusionModel

2. Create a Virtual Environment (Recommended)

It is highly recommended to create a virtual environment to manage dependencies.

python -m venv venv
source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

3. Install Dependencies

Install all the required packages using the requirements.txt file. (Note: we used Tensorflow 2.14)

pip install -r requirements.txt

4. Dataset Preparation

This code relies on the public HAM10000 dataset.

1. Download the dataset: You can find it on Kaggle or other academic sources.

2. Structure the data: The script expects the dataset to be organized into train and val directories, with subdirectories for each of the 7 classes. Your final data structure should look like this:

   /path/to/HAM10000/
   ├── train/
   │   ├── AKIEC/
   │   ├── BCC/
   │   ├── BKL/
   │   ├── DF/
   │   ├── MEL/
   │   ├── NV/
   │   └── VASC/
   └── val/
       ├── AKIEC/
       ├── BCC/
       ├── BKL/
       ├── DF/
       ├── MEL/
       ├── NV/
       └── VASC/
   

3. Update Paths in FusionModel.py: Before running, you must update the hardcoded paths in the FusionModel.py script to point to your dataset location. Look for the variable main_path and other similar path definitions.

---

How to Use the Code

The FusionModel.py script is organized into sections based on the comments (In[0], In[1], etc.). You can run these sections sequentially in an interactive environment like Spyder, VSCode, or a Jupyter Notebook.

1. Model Training

• Ensure your dataset is correctly structured and paths are updated.
• Run the sections In[0] (data loading) and In[1] (model definition).
• Execute section In[2] to start the training process. Model weights (.keras files) and training history (.csv file) will be saved in your working directory.

2. Reporting Results and Evaluation

• After training, load the weights of your best model by updating the model_path variable in section In[3.1].
• Run section In[4] and its subsections to:
  • Generate and display the confusion matrix (In[4.1]).
  • Print and save the classification report (In[4.2]).
  • Compute and plot the ROC curves for each class (In[4.3]).
  • Save predicted probabilities to an Excel file (In[4.4]).

3. Generating Explainable AI (XAI) Visualizations

Grad-CAM

• Load a trained model first (section In[3.1]).
• In section In[3.2], specify the class (skin_class) and image paths (img_paths) you want to visualize.
• Run section In[3.3] to generate and save the Grad-CAM plots to the specified output directory.

SHAP

• Load a trained model first (section In[3.1]).
• Run section In[5.1] to prepare the data for the SHAP explainer.
• In section In[5.3], specify the class (skin_class) and image paths (img_paths) for visualization.
• Run section In[5.4] to generate and save the SHAP value plots.

---

Citation

If you use this code or our work in your research, please cite our paper:

@article{A2025,
  title={Explainable Depth-Wise and Channel-Wise Fusion Models for Multi-Class Skin Lesion Classification},
  author={A, H and S, R and E, M},
  journal={PLOS ONE},
  year={2025}
}

License

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

Acknowledgments

This work utilizes the HAM10000 dataset, collected and made publicly available by Tschandl et al. We thank the authors and contributors for their invaluable contribution to the research community.