FDCN-C: A deep learning model based on frequency enhancement, deformable convolution network, and crop module for...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) electroencephalography (EEG) decoding is crucial for brain-computer interfaces (BCIs), enabling communication for motor-disabled individuals.
• Current deep learning (DL) methods for EEG decoding show limitations in efficiently utilizing temporal and frequency domain features, leading to suboptimal MI classification.
• There is a need for advanced models that can enhance feature extraction from EEG signals for improved BCI performance.

Purpose of the Study:

• To propose a novel EEG-based MI classification model, termed FDCN-C, designed to overcome the limitations of existing methods.
• To enhance the extraction and utilization of frequency and temporal features from EEG signals for more accurate MI classification.
• To validate the effectiveness of the proposed FDCN-C model using public datasets and compare its performance against state-of-the-art methods.

Main Methods:

• The proposed FDCN-C model integrates a frequency enhancement module with attention mechanisms to capture diverse frequency band features.
• A deformable convolutional network is employed for advanced temporal feature extraction by adaptively modulating convolution kernel sizes.
• Spatial information is integrated using a 1D convolution layer, and a crop module with dilated convolutions enhances feature representation across varied receptive fields.

Main Results:

• The FDCN-C model demonstrated superior MI classification accuracy compared to existing state-of-the-art methods on two public datasets.
• The proposed model achieved a significant improvement of 14.01% in accuracy compared to the baseline model.
• Ablation studies confirmed the individual contribution and effectiveness of each component within the FDCN-C model.

Conclusions:

• The FDCN-C model represents a significant advancement in EEG-based MI classification for BCIs.
• The innovative frequency enhancement and deformable convolution modules effectively address the limitations of previous approaches.
• The proposed model offers enhanced performance, paving the way for more robust and efficient brain-computer interfaces.