This study presents a multi-branch convolutional neural network (CNN) designed to improve the classification of motor imagery (MI) EEG signals. MI tasks produce complex EEG patterns with spatial, spectral, and temporal components. The network given in this article extracts these features simultaneously, enabling more accurate discrimination between different MI tasks. Evaluation on benchmark datasets shows that the model outperforms traditional classifiers and standard CNNs. Grad-CAM visualizations highlight the spatial channels and frequency bands most influential in the model’s decisions, providing interpretability and insight into the underlying neural activity. By effectively handling the multidimensional nature of EEG signals and inter-subject variability, this approach enhances the robustness and reliability of motor imagery EEG classification, offering potential improvements for real-world BCI applications. The approach also demonstrates the importance of integrating complementary EEG features, rather than analyzing them in isolation, for achieving high classification accuracy. Its ability to capture subtle neural patterns could inform future research into individualized BCI training protocols. Overall, this method represents a meaningful advance in EEG-based BCI technology, combining performance with interpretability.