Multi-Band CNN With Band-Dependent Kernels and Amalgamated Cross Entropy Loss for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are widely used for motor imagery (MI) classification from electroencephalography (EEG) signals.
• Existing CNN approaches suffer from subject dependency due to kernel size optimization challenges.
• Exploiting frequency diversity in EEG signals is crucial for improving MI classification.

Purpose of the Study:

• To develop a novel MI classification method that resolves subject dependency by optimizing kernel sizes for different frequency bands.
• To introduce a multi-band CNN architecture (MBK-CNN) that leverages frequency diversity in EEG signals.
• To propose an enhanced version (MBK-LR-CNN) with improved spatial diversity for superior classification performance.

Main Methods:

• EEG signals are decomposed into overlapping multi-bands.
• Each band is processed by dedicated CNNs ('branch-CNNs') with unique, band-dependent kernel sizes.
• An amalgamated cross-entropy loss function is employed to prevent overfitting and optimize the network.
• An enhanced spatial diversity approach (MBK-LR-CNN) utilizes sub-branch CNNs on channel subsets ('local regions').

Main Results:

• The proposed MBK-CNN and MBK-LR-CNN methods demonstrate improved MI classification performance.
• Experimental results on public datasets (BCI Competition IV dataset 2a, High Gamma Dataset) confirm performance gains.
• The novel approach effectively addresses the subject dependency issue inherent in previous CNN-based methods.

Conclusions:

• The MBK-CNN architecture successfully exploits EEG frequency diversity and resolves subject dependency for MI classification.
• The MBK-LR-CNN further enhances performance through improved spatial diversity.
• These novel methods represent a significant advancement in EEG-based MI classification accuracy.