Centroid-guided pseudo-label refinement and dual-level contrastive learning for motor-execution-to-motor-imagery EEG adaptation

Electroencephalography (EEG)-based motor imagery (MI) decoding typically relies on labeled MI data for subject-specific calibration, while MI task compliance is difficult to verify objectively, raising concerns about label reliability. In contrast, motor execution (ME) can be directly observed, providing more reliable supervision. However, differences in neural activity patterns and data distributions between ME and MI lead to a cross-task distribution shift and suboptimal direct-transfer performance. To address this issue, we propose Centroid Refinement and Dual-Level Contrastive Learning with a Multi-Scale Depthwise Convolutional Network (CRDCL-MSDCNet), a sequential two-stage framework for subject-specific ME-to-MI adaptation. In Stage 1, labeled ME data from multiple subjects are pooled to train the multi-scale depthwise convolutional network. In Stage 2, subject-specific adaptation uses labeled ME data as the source domain and unlabeled MI data as the target domain. Centroid-Guided pseudo-label refinement iteratively updates class centroids and target assignments, while dual-level contrastive learning enhances the intra-class compactness and inter-class separability of target MI representations through center-level and instance-level constraints. Experimental results show that CRDCL-MSDCNet achieves accuracies of 80.52 ± 11.55% and 79.56 ± 8.52% on EEGMMIDB and the lower-limb dataset, respectively. It outperforms the best compared method (i.e., the Joint Adaptation Network) by 0.43 and 0.67 percentage points on the two datasets, respectively. The ablation study demonstrates that performance was reduced by 3.48 and 7.98 percentage points without Stage 2 adaptation and by 2.40 and 2.20 percentage points when only source cross-entropy (Source CE) was used. These results demonstrate effective transfer of discriminative knowledge from labeled ME to unlabeled MI without using MI labels for training or adaptation.