The article introduces EEG-SGENet, a lightweight, high-performing deep learning model designed to classify motor imagery EEG signals more accurately. While deep learning has improved MI-BCI accuracy, many current models have become very large and computationally expensive. This limits their real-world use, especially in wearable or real-time devices. This is addressed by integrating an attention-based module called Spatial Group-wise Enhance (SGE) into an already efficient CNN architecture (EEGNex). The SGE module reweighs different spatial feature groups so the network pays more attention to important brain regions while suppressing noise from muscle artifacts, electrode noise, and low SNR. Two datasets, BCI Competition IV 2a and 2b, were utilized, and it was shown that the EEG-SGENet achieved 80.98% accuracy on the 2a dataset and 76.17% on the 2b dataset. These results outperform established models such as EEGNet, ShallowConvNet, and the baseline EEGNex. With the combination of multi-scale temporal convolutions, depth-separable spatial convolutions, and SGE-driven attention, the experiments showed improvement with SGE modules. When the authors removed SGE, accuracy dropped consistently across subjects. By demonstrating strong performance even on noisy, variable subject data, EEG-SGENet suggests improved reliability for patient populations who cannot maintain strict stillness or precise electrode conditions. The ability to enhance physiologically meaningful spatial patterns may also improve how BCIs adapt to neural reorganization during rehabilitation and, therefore, represents a reliable BCI to help neurorehabilitation.