Objective In this paper, we propose Ghost-LENet, a lightweight convolutional neural network based on the LENet framework for electroencephalogram (EEG)-based brain–computer interfaces (BCIs), particularly for motor imagery (MI) classification tasks.Methods The proposed model integrates several lightweight design components to improve feature representation while maintaining a very small parameter scale. Specifically, dilated temporal convolutions and stationary wavelet transform are combined in the initial block to capture multi-scale temporal characteristics of EEG signals. In addition, a dynamic residual fusion mechanism with Efficient Channel Attention (ECA), referred to as DR-ECA, is introduced to adaptively balance attention-enhanced features and original features. Furthermore, a Ghost module is incorporated to improve parameter efficiency while preserving feature extraction capability.Results Experimental results on multiple public EEG datasets show that Ghost-LENet achieves competitive classification performance with only a few thousand trainable parameters across datasets. Specifically, Ghost-LENet achieves classification accuracies of 82.18% and 83.05% on the BCI Competition IV-2a and IV-2b datasets, respectively.Conclusion Overall, Ghost-LENet provides a lightweight and efficient decoding framework for EEG-based motor imagery classification, showing a favorable balance between model complexity and classification performance.