Enhancing-Cross-Subject-Motor-Imagery-Classification-using-VGG-based-CNN-and-WGANs

This repository contains the code and data for the paper "Toward Calibration-Free Motor Imagery Brain-Computer Interfaces: A VGG-based Convolutional Neural Network and WGAN Approach"by A. G. Habashi, Ahmed M. Azab, Seif Eldawlatly, and Gamal M. Aly. The paper proposes a novel framework for cross-subject motor imagery (MI) classification using electroencephalogram (EEG) signals. The framework employs a Wasserstein Generative Adversarial Network (WGAN) for data augmentation and a modified VGG-based Convolutional Neural Network (CNN) for classification.

Data

The code is implemented using one benchmark dataset (BCI Competition IV-2B) as an example. However, the evaluation in the paper was conducted on three benchmark datasets (BCI Competition IV-2B, IV-1, and IV-2A). The data can be downloaded from the BCI Competition IV website: https://www.bbci.de/competition/iv/

Preprocessing

The preprocessing steps involve filtering the EEG data, extracting MI task segments, computing Short-Time Fourier Transform (STFT) spectrograms, and merging and resizing the spectrograms into grayscale images. The preprocessing scripts for each dataset are provided in the repository:

preprocessing dataset 2B.py
preprocessing dataset 1.py
preprocessing dataset 2A.py

WGAN Training

The WGAN is trained to generate synthetic EEG spectrum images that are similar to the real data. The WGAN architecture and training process are implemented in the func_cnn_Wgan.py script.

CNN Training and Testing

The modified VGG-based CNN classifier is trained on the augmented dataset (real and synthetic images). The training and testing process is implemented in the Gray cross-WGAN-CNN.py script. The script also includes the implementation of two other CNN classifiers (VGG-based CNN and Audio-spectrum CNN) for comparison.

Evaluation

The framework is evaluated using a leave-one-subject-out cross-validation scheme. The classification accuracy is used as the performance metric. The results demonstrate that the proposed framework outperforms state-of-the-art methods in cross-subject MI classification.

Requirements

Python 3.9.7
TensorFlow
Keras
NumPy
SciPy
Matplotlib
gumpy

Usage

Download the datasets from the BCI Competition IV website.
Preprocess the datasets using the provided scripts.
Train the WGAN using the func_cnn_Wgan.py script.
Train and test the CNN classifiers using the Gray cross-WGAN-CNN.py script.

Note

The code provided in this repository is for research purposes only. It may require modifications to be used in real-world applications.

Ahmed-Habashy/Enhancing-Cross-Subject-Motor-Imagery-Classification-using-VGG-based-CNN-and-WGANs