Attention Induced Dual Convolutional-Capsule Network (AIDC-CN): A deep learning framework for motor imagery...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electroencephalography (EEG)-based motor imagery (MI) decoding is crucial for assistive technologies but faces challenges due to signal complexity and noise.
• Deep learning models are increasingly used for EEG signal decoding, yet advancements are needed for improved accuracy.

Purpose of the Study:

• To introduce a novel deep learning classifier, the Attention Induced Dual Convolutional-Capsule Network (AIDC-CN), for accurate motor imagery classification.
• To enhance EEG decoding performance by employing a dual feature extraction approach using spectrograms and brain connectivity networks.

Main Methods:

• Developed the Attention Induced Dual Convolutional-Capsule Network (AIDC-CN) incorporating dual convolution layers, a self-attention module (SAM), a cross-attention module (CAM), and a GELU-based dynamic routing algorithm.
• Utilized a dual feature extraction strategy combining spectrograms and brain connectivity networks to enrich the feature set for classification.
• Evaluated the AIDC-CN model on four publicly available datasets.

Main Results:

• The proposed AIDC-CN model achieved superior performance compared to state-of-the-art techniques across four public datasets.
• The integration of dual feature extraction, attention mechanisms, and capsule networks significantly improved motor imagery classification accuracy.
• The novel architectural components, including SAM, CAM, and GELU-based routing, effectively addressed the challenges of EEG signal decoding.

Conclusions:

• The AIDC-CN represents a significant advancement in deep learning for EEG-based motor imagery decoding.
• The model's effectiveness in accurately classifying motor imagery tasks highlights its potential for applications in assistive robotics and neurorehabilitation.
• The study provides a robust and high-performing deep learning framework for complex EEG signal analysis.