BCIwiki (bciwiki.com) — A lightweight convolutional neural network called Ghost-LENet achieved 82.18% and 83.05% classification accuracy on the BCI Competition IV-2a and IV-2b datasets, respectively, for EEG motor imagery classification, with only a few thousand trainable parameters. The study, by researchers at Northwest Normal University, Fudan University, and other institutions, was published in Brain-Apparatus Communication: A Journal of Bacomics on July 20, 2026.
Ghost-LENet, built on the LENet framework, integrates dilated temporal convolutions and stationary wavelet transform to capture multi-scale temporal features of EEG signals, introduces a dynamic residual fusion mechanism (DR-ECA) to adaptively balance attention-enhanced and original features, and employs a Ghost module to improve parameter efficiency. Experimental results across multiple public EEG datasets show that the model achieves competitive classification performance with a very small parameter scale, offering a lightweight and efficient decoding framework for resource-constrained BCI applications.