摘要
Children with developmental coordination disorder (DCD) often experience difficulties with motor imagery (MI), the mental rehearsal of movement. This limitation presents an opportunity in which MI-based brain-computer interface (BCI) systems, which detect neural correlates of imagined movement, may serve as diagnostic tools for identifying motor planning deficits in children with DCD. However, the absence of behavioral confirmation during MI tasks often results in noisy EEG labels that complicate the development of accurate models. To address this challenge, we propose DualRefine, a novel semi-supervised dual-network framework designed to enhance model robustness in the presence of label noise. Built upon an EEGNet backbone, DualRefine consists of two peer networks with distinct classification heads that collaboratively perform data categorization, pseudo-labeling, and iterative model refinement through mutual supervision. Compared to existing learning-with-noisy-label (LNL) approaches, DualRefine introduces three major innovations: (1) a co-learning strategy that cross-validates peer models, (2) confidence-based filtering to isolate reliable data, and (3) a relabeling mechanism to recover mislabeled samples. Evaluated on EEG datasets from MI tasks, DualRefine achieved 71.6 +/- 1.3% accuracy, significantly outperforming other LNL methods (approximately 50%). For DCD classification, it reached 76.0 +/- 4.9%, markedly improving upon the 63% accuracy of baseline models. These results demonstrate the general applicability of DualRefine beyond a single dataset or condition, offering a scalable solution for EEG-based classification tasks where ground truth is limited or unreliable. This work is the first to explicitly integrate noisy-label countermeasures into EEGbased MI-BCI pipelines, paving the way for more robust BCI applications in both clinical and broader neurotechnology domains where unreliable EEG labels remain a key challenge.