A multiscale siamese convolutional neural network with cross-channel fusion for motor imagery decoding

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are prevalent in motor imagery electroencephalography (MI-EEG) signal classification.
• Simple CNN frameworks struggle with the intricate decoding requirements of MI-EEG signals.

Purpose of the Study:

• To introduce a novel deep learning architecture, MSCCF-Net, for improved MI-EEG signal classification.
• To enhance the representation of multiscale temporal features and optimize the training process for complex EEG data.

Main Methods:

• Proposed a multiscale Siamese convolutional neural network with cross-channel fusion (MSCCF-Net).
• Incorporated Siamese cross-channel fusion streams with multiple branches and cross-channel fusion modules for enhanced feature extraction.
• Utilized a similarity module for feature comparison and a classification module for robust classification, employing a joint training strategy.

Main Results:

• The MSCCF-Net was evaluated on the BCI Competition IV 2a and 2b datasets.
• Achieved an average accuracy of 87.36% on the BCI Competition IV 2a dataset.
• Achieved an average accuracy of 87.33% on the BCI Competition IV 2b dataset.

Conclusions:

• The MSCCF-Net effectively utilizes cross-channel fusion for multiscale temporal feature learning.
• The joint training strategy optimizes the network's performance in MI-EEG classification.
• The proposed method demonstrates superior performance compared to existing state-of-the-art MI-EEG classification techniques.