Enhancing EEG Artifact Removal Efficiency by Introducing Dense Skip Connections to IC-U-Net

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electroencephalographic (EEG) data is often corrupted by artifacts, necessitating effective removal techniques.
• Deep learning models, like the proposed IC-U-Net, have shown promise in enhancing EEG signal-to-noise ratio (SNR) and brain-computer interface (BCI) performance.
• Existing models face challenges with overfitting on limited data and high computational costs.

Purpose of the Study:

• To enhance the practicability of the IC-U-Net model for EEG artifact removal.
• To address overfitting and high computational demands of previous deep learning approaches.
• To improve EEG signal quality for better BCI applications.

Main Methods:

• Leveraged the UNet++ architecture, incorporating dense skip connections into the encoder-decoder framework.
• Modified the existing IC-U-Net model to create a more efficient and robust EEG artifact removal system.
• Evaluated the model's performance in terms of SNR improvement, parameter count, and convergence speed.

Main Results:

• The proposed UNet++ based model achieved a superior signal-to-noise ratio (SNR) compared to the original IC-U-Net model.
• The new model demonstrated this improvement with half the number of parameters.
• Comparable convergence was achieved using only a quarter of the training data size, indicating enhanced efficiency.

Conclusions:

• The UNet++ architecture effectively improves EEG artifact removal, offering better practicability than previous models.
• This enhanced model reduces computational cost and data requirements, making it more suitable for real-world applications.
• The findings suggest a significant advancement in deep learning-based EEG signal processing for improved BCI performance.