Msst-eegnet: multi-scale spatio-temporal feature extraction using inception and temporal pyramid pooling for motor...

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interface (BCI) Systems
• Deep Learning for Signal Processing

Background:

• Motor imagery classification is crucial for interpreting brain signals in BCI systems.
• Accurate classification enables intuitive control of external devices via thought.
• Existing deep learning models face challenges in extracting comprehensive features for robust classification.

Purpose of the Study:

• To develop a novel deep learning (DL) model for enhanced motor imagery classification.
• To extract discriminative spatio-temporal and spectral features for improved generalization.
• To achieve superior performance compared to existing state-of-the-art methods.

Main Methods:

• Introduction of the Multi-scale spatio-temporal network (MSST-EEGNet).
• Utilized an inception module with dilated convolution for multi-scale temporal and spatial feature extraction.
• Employed a temporal pyramid pooling module for fine-grained and coarse-grained feature extraction.
• Implemented categorical cross-entropy combined with center loss as the optimization objective.

Main Results:

• MSST-EEGNet demonstrated superior classification accuracy on three benchmark datasets (BCI Competition IV-2a, IV-2b, OpenBMI).
• Outperformed eight existing deep learning models in subject-specific and cross-session classifications.
• Outperformed eight deep learning and six transfer-learning models in cross-subject classification.
• Achieved high accuracies, e.g., 0.8426 ± 0.1061 for subject-specific on BCI IV-2a.

Conclusions:

• The proposed MSST-EEGNet model offers a significant advancement in motor imagery classification.
• The network effectively extracts multi-scale features, leading to enhanced generalization.
• Statistical tests confirm the superior performance of MSST-EEGNet over existing approaches.