Decoding 3D Representation of Visual Imagery EEG using Attention-based Dual-Stream Convolutional Neural Network

TL;DR

This paper introduces an attention-based dual-stream 3D CNN that improves EEG-based visual imagery classification by better capturing spatial inter-region relationships, achieving higher accuracy and revealing neurophysiological insights.

Contribution

It presents a novel dual-stream 3D CNN with channel attention for enhanced spatial feature extraction in EEG-based visual imagery classification.

Findings

Achieved 0.58 accuracy in 4-class visual imagery EEG classification.

Visual motion imagery has higher alpha-power spectral density over the visual cortex.

Swarm dispersion imagery shows higher beta-PSD over pre-frontal cortex.

Abstract

A deep neural network has been successfully applied to an electroencephalogram (EEG)-based brain-computer interface. However, in most studies, the correlation between EEG channels and inter-region relationships are not well utilized, resulting in sub-optimized spatial feature extraction. In this study, we propose an attention-based dual-stream 3D-convolutional neural network that can enhance spatial feature extraction by emphasizing the relationship between channels with dot product-based channel attention and 3D convolution. The proposed method showed superior performance than the comparative models by achieving an accuracy of 0.58 for 4-class visual imagery (VI) EEG classification. Through statistical and neurophysiological analysis, visual motion imagery showed higher alpha-power spectral density (PSD) over the visual cortex than static VI. Also, the VI of swarm dispersion showed higher beta-PSD over the pre-frontal cortex than the VI of swarm aggregation.