北京生物医学工程2025，Vol.44Issue(3)：237-244,8.DOI:10.3969/j.issn.1002-3208.2025.03.003
基于双注意力的多尺度卷积神经网络
A dual attention based multi-scale convolution neural network
摘要
Abstract
Objective It is a research hotspot in neurorehabilitation to recognize motor-imagery electroencephalogram(MI-EEG)signals based on convolutional neural networks(CNNs).However,the current CNN models fail to simultaneously explore the rich temporal-spectral-spatio features and effectively address the personalized characteristics of MI-EEG.Therefore,we propose a dual attention based multi-scale CNN(DAMSCNN)to realize multi-domain feature extraction and classification for motor imagery.Methods First,a filter bank is employed to filter the original motor imagery signal.Then,the multi-band data is spliced to construct three-dimensional MI-EEG data in the temporal-spectral-spatio domains,which is input into a dual-attention module.The squeeze-excitation block(SE Block)and self-attention block(SA Block)are utilized to compute attention weights across different spectral bands and electrode channels,respectively,thereby enhancing the focus on informative spectral and spatial features.Then,the multi-scale temporal convolution module is applied in parallel to extract multi-scale temporal features from key spectral-spatial information,which is then integrated with the three-dimensional MI-EEG data.The spatial-spectral convolution module is used to extract and fuse all features from all frequency bands and channels.Finally,the classification module completes the MI classification.To verify the effectiveness and feasibility,ablation experiments and comparative experiments are conducted on the public BCI Competition IV 2a and accuracy and F1 value are used as evaluation indicators.Results DAMSCNN achieves the average classification accuracies of 84.02%and 79.81%,and F1 values of 82.32%and 79.67%in within-session and cross-session,respectively.Conclusions The results show that DAMSCNN can adaptively capture the temporal-spectral-spatio features of MI-EEG.It enhances the attention importance of useful information and improves the accuracy of motor imagery classification.关键词
运动想象信号/多尺度卷积/自注意力机制/挤压-激励机制/特征融合Key words
motor Imagery/multi-scale convolution/self-attention/squeeze and exication/data augmentation分类
基础医学引用本文复制引用
刘赛楠,李明爱..基于双注意力的多尺度卷积神经网络[J].北京生物医学工程,2025,44(3):237-244,8.基金项目
国家自然科学基金(62173010)资助 （62173010）
