Generative Adversarial Network with Adaptive Synthesis for Brain-Computer Interfaces in Motor Imagery Classification

Abstract

Motor Imagery (MI) is essential in Brain-Computer Interfaces (BCIs), highlighting the central role of electroencephalography (EEG) in this technology. However, the amount of raw EEG data is often limited. Raw EEG data contains significant noise caused by individual and task-specific differences. These complications often necessitate the removal of noise, making it crucial to develop robust models capable of maintaining performance. To achieve this, synthetic data is essential to obtain cleaner signals. This paper presents a novel a novel contrastive diffusion model to adaptive synthesize clean signals for Motor Imagery Classification (CDASC-MI), achieved through three components: Clean Signal Extractor(CSE) for clean signal feature learning, Adapative Noise Extractor(ANE) for adaptively extracting subject noise and task noise, and Contrastive Module for further separating noise to obtain cleaner signals. Specifically, we first proposed the novel Cell-UNET in CSE, which is an architecture with progressive learning capabilities to effectively capture subtle changes in EEG signals. Additionally, ANE incorporates a novel Subject Feature Learning Network(SFLNet) and Cross-Domain Electrode Attention Network (CDANet). SFLNet and CDANet dynamically capturing variation subject noise and task noise. Ultimately, a Contrastive Module is designed to use contrastive learning during the denoising process to amplify the differences between clean signals and noise, ensuring robust denoising and clearer feature representation. According to experimental results from two public datasets, the proposed method achieved superior accuracies of 96.19% and 89.19%, outperforming existing approaches. It highlights its potential as a powerful method for augmenting EEG datasets and improving the accuracy and highlight its broad application values.