EEG-DCNet: A Fast and Accurate MI-EEG Dilated CNN Classification Method

Abstract: The electroencephalography (EEG)-based motor imagery (MI) classification is a critical and challenging task in brain-computer interface (BCI) technology, which plays a significant role in assisting patients with functional impairments to regain mobility. We present a novel multi-scale atrous convolutional neural network (CNN) model called EEG-dilated convolution network (DCNet) to enhance the accuracy and efficiency of the EEG-based MI classification tasks. We incorporate the $1\times1$ convolutional layer and utilize the multi-branch parallel atrous convolutional architecture in EEG-DCNet to capture the highly nonlinear characteristics and multi-scale features of the EEG signals. Moreover, we utilize the sliding window to enhance the temporal consistency and utilize the attension mechanism to improve the accuracy of recognizing user intentions. The experimental results (via the BCI-IV-2a ,BCI-IV-2b and the High-Gamma datasets) show that EEG-DCNet outperforms existing state-of-the-art (SOTA) approaches in terms of classification accuracy and Kappa scores. Furthermore, since EEG-DCNet requires less number of parameters, the training efficiency and memory consumption are also improved. The experiment code is open-sourced at \href{https://github.com/Kanyooo/EEG-DCNet}{here}.

• The paper demonstrates that EEG-DCNet significantly improves motor imagery EEG classification by integrating multi-scale atrous CNN blocks for robust feature extraction.
• The methodology employs efficient 1x1 convolutions and parallel atrous branches to reduce parameters while optimizing accuracy on benchmark BCI datasets.
• Experimental results show improved accuracy (87.94%) and Kappa score (0.8392) compared to state-of-the-art models, highlighting its potential for real-time BCI applications.

EEG-DCNet: A Fast and Accurate MI-EEG Dilated CNN Classification Method

The research paper presents the development of EEG-DCNet, a novel model designed to improve the classification of motor imagery (MI) electroencephalography (EEG) signals, which are vital for enhancing brain-computer interface (BCI) applications. EEG signals provide a non-invasive means to capture brain activity with high temporal resolution but face challenges related to low signal fidelity, individual variability, and inadequate feature extraction. EEG-DCNet addresses these issues through a multi-scale atrous convolutional neural network (CNN) architecture that incorporates several innovations.

Key Contributions

1. Multi-scale Atrous Convolution Block: The EEG-DCNet incorporates a multi-scale atrous CNN block to capture non-linear characteristics and diverse scale features inherent in EEG signals. By replacing average pooling with this block, the model enhances representation capabilities and reduces information loss during signal transmission.
2. Efficient Architectural Components: The integration of $1 \times 1$ convolutional layers and multi-branch parallel atrous convolutions allows for the extraction and effective integration of multi-scale information, which improves the generalization of the model across datasets. This configuration supports capturing continual changes in EEG signals, essential for maintaining accuracy.
3. Temporal Consistency through Sliding Window and Attention Mechanisms: EEG-DCNet employs a sliding window method coupled with attention mechanisms to boost temporal consistency. This approach ensures the network captures the important temporal segments of EEG data, thereby improving the recognition of user intent during MI tasks.

Experimental Validation

The model was evaluated on multiple datasets, including BCI-IV-2a, BCI-IV-2b, and the High-Gamma dataset. In these tests, EEG-DCNet demonstrated superior performance in accuracy and Kappa scores compared to existing state-of-the-art methods. On the BCI-IV-2a dataset, EEG-DCNet achieved an accuracy of 87.94% and a Kappa score of 0.8392, outperforming all baseline models tested, including EEGNet, ShallowConvNet, MBEEG_SENet, EEGTCNet, EEGNeX, and ATCNet. Additionally, it requires fewer parameters, leading to improved training efficiency and reduced computational costs.

Implications and Future Directions

The architectural advancements of EEG-DCNet underscore its utility in both theoretical and practical realms for EEG classification. The proposed framework not only enhances feature extraction but also provides a scalable solution that can accommodate individual variability—a noteworthy achievement given the diverse nature of EEG data.

However, the research identifies the need for improvements, particularly in the context of generalization and real-time application efficiency. Abilities like incremental learning remain unexplored within this framework but hold significant promise for continuous system adaptation in real-world settings. Future investigations could focus on deploying EEG-DCNet on edge devices with strategies like transfer learning and real-time update capabilities.

Attention mechanisms leveraged in EEG-DCNet present another avenue for understanding critical EEG signal features that influence classification decisions, fostering both clinical acceptance and user trust. Additionally, the integration of multimodal data could potentially enhance performance by reducing individual variability effects on EEG signals.

In summary, EEG-DCNet represents a significant advancement in MI EEG classification, balancing high performance with efficient computation, and paving the way for future exploration and enhancements in EEG-based BCI technologies.