Scholarlink

<p>Motor imagery (MI) is currently one of the most researched brain‒computer interface (BCI) paradigms, with convolutional neural networks (CNNs) being extensively used for decoding electroencephalogram (EEG) signals. However, numerous studies have demonstrated that the optimal convolution scale varies across subjects and even within different sessions for the same subject. Additionally, EEG signals are time-series data in which temporal dependencies are critical for decoding performance. Consequently, CNN-based models, which are limited by their receptive fields, often struggle to extract highly discriminative features from EEG signals with low signal-to-noise ratios. To address these challenges, an improved multi-scale convolution and Transformer network (IMCTNet) is proposed for decoding MI tasks. In this network, we design a multi-scale temporal convolution block with a channel attention mechanism, enabling adaptive weighting of feature maps across scales. Additionally, we employ Transformer to capture global temporal dependencies, compensating for the limited receptive field of CNNs. The experimental results demonstrate that the IMCTNet outperforms state-of-the-art models, achieving superior classification accuracy (81.83% for Dataset BCI competition IV 2a, 86.47% for Dataset BCI competition IV 2b, and 74.19% for Dataset OpenBMI). The code for IMCTNet is available at <a href="https://github.com/ydxswys1/EEG-IMCTNet">https://github.com/ydxswys1/EEG-IMCTNet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An improved multi-scale convolution and transformer network for EEG-based motor imagery decoding

  • 期刊论文

  • 发表日期：

    2025年2月4日

查看全文
  • Lei Zhu,
  • Yunsheng Wang,
  • Aiai Huang,
  • Xufei Tan,
  • Jianhai Zhang

摘要

Motor imagery (MI) is currently one of the most researched brain‒computer interface (BCI) paradigms, with convolutional neural networks (CNNs) being extensively used for decoding electroencephalogram (EEG) signals. However, numerous studies have demonstrated that the optimal convolution scale varies across subjects and even within different sessions for the same subject. Additionally, EEG signals are time-series data in which temporal dependencies are critical for decoding performance. Consequently, CNN-based models, which are limited by their receptive fields, often struggle to extract highly discriminative features from EEG signals with low signal-to-noise ratios. To address these challenges, an improved multi-scale convolution and Transformer network (IMCTNet) is proposed for decoding MI tasks. In this network, we design a multi-scale temporal convolution block with a channel attention mechanism, enabling adaptive weighting of feature maps across scales. Additionally, we employ Transformer to capture global temporal dependencies, compensating for the limited receptive field of CNNs. The experimental results demonstrate that the IMCTNet outperforms state-of-the-art models, achieving superior classification accuracy (81.83% for Dataset BCI competition IV 2a, 86.47% for Dataset BCI competition IV 2b, and 74.19% for Dataset OpenBMI). The code for IMCTNet is available at https://github.com/ydxswys1/EEG-IMCTNet.