Decoding motor imagery from non-invasive EEG remains challenging due to the low signal-to-noise ratio and scarce labeled data across subjects. We present MINDGAN, a unified framework that integrates class-conditional deep convolutional GAN (cDCGAN) augmentation with a hybrid CNN–Transformer classifier under a three-phase curriculum training scheme supported by a quality-filtered replay buffer. A systematic six-configuration ablation reveals that recurrent components (LSTM) are detrimental within hybrid architectures. MINDGAN achieves competitive performance — 81.17% accuracy on BCI Competition IV Dataset 2A and 87.16% on Dataset 2B — with the lowest cross-subject variance on 2A.
@inproceedings{ahmad2026mindgan,title={MINDGAN: EEG-Based Motor Imagery Decoding via Hybrid CNN--Transformer with Curriculum GAN Augmentation},author={Ahmad, Meshkat and Huang, Yanqi and Wu, Xiaomei},booktitle={2026 9th IEEE International Conference on Communication Engineering and Technology (ICCET)},year={2026},month=jun,address={Guangzhou, China},pages={18--24},publisher={IEEE},doi={10.1109/ICCET70051.2026.11659043},url={https://doi.org/10.1109/ICCET70051.2026.11659043},}