Motor imagery classification is an essential component of Brain-computer interface systems
to interpret and recognize brain signals generated during the visualization of motor imagery
tasks by a subject. The objective of this work is to develop a novel DL model to extract
discriminative features for better generalization performance to recognize motor imagery
tasks. This paper presents a novel Multi-scale spatio-temporal network (MSST-EEGNet)
to extract discriminative temporal, spectral, and spatial features for motor imagery task
classification. The proposed MSST-EEGNet model includes three modules namely the
inception module with dilated convolution, the temporal pyramid pooling module, and
the classification module. Multi-scale temporal features along with spatial features are
extracted using the inception block with the dilated convolution module. A set of multi-level
fine-grained and coarse-grained features are extracted using a temporal pyramid pooling
module. Further, categorical cross-entropy in combination with center loss is used as a
loss function. Experiments are carried out on three benchmark datasets including the BCI
Competition IV-2a dataset, the BCI Competition IV-2b dataset, and the OpenBMI dataset.
The evaluation results shows that the proposed MSST-EEGNet model outperforms eight
existing DL models in terms of classification accuracy for subject-specific and cross-session
settings. It also outperforms eight existing DL models and six existing transfer-learning
models for cross-subject setting. For the subject-specific classification the proposed MSST-EEGNet model achieved an accuracy of 0.8426 ± 0.1061, 0.7779 ± 0.0938, and 0.7365 ±
0.1477 on the BCI Competition IV-2a dataset, the BCI Competition IV-2b dataset, and the
OpenBMI dataset respectively. For the cross-session setting, the proposed MSST-EEGNet
model achieved an accuracy of 0.7709 ± 0.1098, 0.7524 ± 0.1017, and 0.6860 ± 0.0990 on the BCI Competition IV-2a dataset, the BCI Competition IV-2b dataset, and the OpenBMI
dataset respectively. For the cross-subject setting, the proposed MSST-EEGNet model
achieved an accuracy of 0.7288 ± 0.0730, 0.8161 ± 0.963, and 0.7075 ± 0.0746 on the
BCI Competition IV-2a dataset, the BCI Competition IV-2b dataset, and the OpenBMI
dataset respectively. Furthermore, a non-parametric Friedman statistical test demonstrates
statistically significant superior performance of the proposed MSST-EEGNet model over
the existing models.