MI-EEGNET: A novel convolutional neural network for motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) enable human-machine interaction via brainwave decoding.
• Convolutional neural networks (ConvNets) advance motor imagery decoding, but shallow models often outperform deep ones.
• Existing deep ConvNets face performance limitations, necessitating novel architectures.

Purpose of the Study:

• To design a novel, deeper ConvNet for motor imagery decoding with improved performance and optimal complexity.
• To overcome the performance limitations of existing deep ConvNet models in BCI applications.

Main Methods:

• Developed a novel ConvNet architecture integrating Inception and Xception principles.
• Employed separable and depthwise convolutions for enhanced efficiency and speed.
• Introduced a new Inception-inspired block to capture richer features for improved classification.

Main Results:

• The proposed ConvNet achieved performance comparable to state-of-the-art techniques.
• Convolutional layer weights provided insights into learned features, highlighting the most relevant ones.
• The model significantly outperformed Filter Bank Common Spatial Pattern (FBCSP), Riemannian Geometry (RG), and ShallowConvNet (p < 0.05).

Conclusions:

• Motor imagery decoding is achievable using deep learning models without relying on handcrafted features.
• The novel ConvNet architecture demonstrates superior performance in BCI applications.
• The findings suggest a promising direction for developing more effective BCI systems.