DecodeMI-Net: A Multi-Perspective Deep Neural Network With Channel-Wise Attention for Motor Imagery EEG Decoding

Abstract

Identification of Motor Imagery (MI) tasks from electroencephalography (EEG) signals plays an important role in brain–computer interfaces (BCIs), especially in the development of non-invasive assistive technologies for individuals with motor impairments. Despite recent advancements, many existing MI decoding methods analyze EEG signals from a limited perspective and struggle to generalize across different subjects and recording sessions. To address these limitations, this paper proposes a multi-branch deep neural network for decoding MI tasks, namely DecodeMI-Net, which processes EEG signals from multiple perspectives. The proposed method integrates spatial and temporal convolutions at multiple resolutions and combinations to achieve enriched and comprehensive feature extraction. It also incorporates a Time-Squeezer module followed by a lightweight channel attention mechanism to capture inter-electrode interactions and temporal variability while keeping computational cost and parameter count minimal. By jointly optimizing each branch toward the common goal of decoding, DecodeMI-Net supports both group-level and subject-specific learning. A topographical image analysis is performed to demonstrate the capability of the proposed model in capturing the variation of neurophysiological activities due to different MI tasks. The proposed method is evaluated on three open-source EEG datasets—BCIC-IV 2a, BCIC-IV 2b, and High-Gamma. On BCIC-IV 2a, it achieves 85.12% accuracy for session-dependent tasks, 83.05% for session-independent tasks, and 70.29% for subject-independent tasks, outperforming state-of-the-art baselines.