---
title: "Deep multiscale CNN with temporal-channel attention for motor imagery EEG decoding"
id: "plos-one-21-decoding-motor-imagery-eeg-signals-by-a-deep-learning-network-integrating"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-21-decoding-motor-imagery-eeg-signals-by-a-deep-learning-network-integrating"
content_type: "clinical_feed_article"
specialty: "Neurology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361026"
published_at: "2026-10-23T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Deep multiscale CNN with temporal-channel attention for motor imagery EEG decoding
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-21-decoding-motor-imagery-eeg-signals-by-a-deep-learning-network-integrating
- **Specialty:** [Neurology](https://medichelpline.com/clinical-feed/neurology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361026)
- **Published At:** 2026-10-23T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- Investigators proposed a deep learning architecture combining a **multiscale CNN** front end with temporal and channel attention modules to decode **motor imagery** EEG. - The network uses four parallel convolutional branches to extract temporal and spatial features, each branch followed by a Transformer block to learn long-range time correlations. - Outputs from the four Transformer blocks are concatenated along the channel axis and passed through a channel attention module for channel fusion before classification. - The model replaces the conventional fully connected classifier with a compact CNN-based classifier to reduce computational complexity. - Training data were augmented by segmentation and reconstruction; the authors reported evaluation on three publicly available BCI datasets (links provided in the paper) rather than new proprietary data. - Reported mean accuracies were 80.21% (Dataset 1, 4-class), 90.19% (Dataset 2, 2-class), and 97.32% (Dataset 3, 4-class), exceeding the compared baselines by modest margins. - The study compared performance against five state-of-the-art deep networks including FBCNet and MSFBCNN, and observed incremental gains on the chosen datasets. - Authors noted challenges typical for MI-BCI work: low EEG signal-to-noise ratio, nonstationarity, and limited training samples; the paper did not report external prospective validation or real-time BCI deployment results. - Methodological specifics such as precise architecture hyperparameters, per-dataset preprocessing steps, and full training schedules are provided in the source article (see Methods); where absent, this summary does not invent them. - The study suggests the approach as a potential method for MI-BCI applications but leaves open questions about generalizability across subjects, cross-session robustness, and hardware deployment constraints.
## Clinical Analysis & Structured Key Points
Decoding motor imagery EEG signals by a deep learning network integrating multiscale filtering and temporal-channel attention mechanisms | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Due to their low signal-to-noise ratio and non-stationarity, accurately classifying EEG signals is still one of the main hindrances that limit the real-world applications of a brain-computer interface (BCI). Deep learning (DL) is a potential method for addressing the problem. However, how to create a suitable DL network for motor imagery (MI)-based BCIs poses severe challenges in the case of insufficient training data. In this paper, a DL network based on multiscale convolutional neural network (CNN) and temporal and channel attention mechanisms is proposed for decoding MI tasks. Temporal and spatial features are extracted by four parallel CNN branches, each of which is followed by a Transformer block for learning the global correlations of local time features. The outputs from the four Transformer blocks are concatenated along channel dimension and then input into a channel attention module for channel fusion. The subsequent feature map is classified by a CNN, which replaces the commonly used fully connected layer for reducing computational complexity. The proposed network was evaluated on three publicly available BCI datasets and compared with five state-of-the-art DL networks. The training data sets are augmented using the approach segmentation and reconstruction. The experimental results demonstrate that the proposed network achieved an average accuracy of 80.21% on Dataset 1 (4-class tasks), 90.19% on Dataset 2 (2-class tasks), and 97.32% on Dataset 3 (4-class tasks), surpassing the strongest baselines FBCNet on Dataset 1, MSFBCNN on Dataset 2 and MSFBCNN on Dataset 3 by 2.59%, 1.27%, and 1.99% respectively. The network effectively decoded the MI-based EEG signals, providing a potential method for BCI applications. Citation: Zhu Y, Shi Y, Wu J, Wei Q (2026) Decoding motor imagery EEG signals by a deep learning network integrating multiscale filtering and temporal-channel attention mechanisms. PLoS One 21(10): e0361026. https://doi.org/10.1371/journal.pone.0361026 Editor: Yuvaraj Rajamanickam, Nanyang Technological University, SINGAPORE Received: May 28, 2026; Accepted: October 5, 2026; Published: October 23, 2026 Copyright: © 2026 Zhu et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: No data was generated by this study. The three publicly available datasets used in the study can be downloaded from the following two websites: http://www.bbci.de/competition/iv/ , https://gin.g-node.org/robintibor/high-gamma-dataset/src/master/data/ . Funding: This work was supported by the National Natural Science Foundation of China (#62066028). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. 1. Introduction A brain-computer interface (BCI) is an intelligent system that enables direct communication between the brain and a computer or other external devices [ 1 ]. The system collects brain electrical signals, decodes these signals and converts them into the control commands understandable by computers. In the past few decades, BCIs have developed rapidly, and BCI-related technologies have been widely applied in medical, entertainment, neural rehabilitation, and other fields [ 2 – 5 ]. Since electroencephalogram (EEG) has the advantages of noninvasiveness, cost efficiency and high time resolution, EEG-based BCIs have become one of the most vibrant areas of research. Motor Imagery (MI) is the mental rehearsal of movement execution without any physical movement or the need of external stimuli. Thereby, an MI-based BCI (MI-BCI) is one kind of most intuitive and natural BCI paradigm. When a subject performs an MI task such as imagining movement of left hand or right hand, the contralateral primary sensorimotor cortex exhibits a decrease in amplitude of the (8–13 Hz) and (14–30 Hz) rhythms, referred to as event-related desynchronization (ERD) [ 6 , 7 ]. The ipsilateral primary sensorimotor cortex, on the other hand, shows an increase in amplitude of the same rhythms, termed event-related synchronization (ERS) [ 8 , 9 ]. This differentiation in rhythmic amplitude can be leveraged to distinguish different MI tasks and thus constitutes the basis of creating an MI-BCI system. Since EEG signals are nonstationary and possess low signal-to-noise ratio, one of main challenges faced by MI-BCIs in real-world applications is to accurately recognize different MI tasks. EEG signal decoding can be performed by either traditional machine learning (ML) or deep learning (DL) methods. As a typical pattern recognition system, a BCI includes two main signal processing steps, namely feature extraction and pattern classification. The traditional ML algorithms for feature extraction include Common Spatial Patterns (CSP) [ 10 , 11 ], Power Spectral Density (PSD) analysis [ 12 ], Short-Time Fourier Transform (STFT) [ 13 ], Wavelet Packet Decomposition (WPD) [ 14 ], Autoregressive Model (AR) [ 15 ], etc. Among them, CSP is an effective spatial filtering and feature extraction method, which can enhance the distinction between two-class MI signals by capturing spatial patterns in EEG signals. So far, CSP has achieved significant success in the MI-BCI field, leading to various CSP variants, the most representative of which is the filter bank CSP (FBCSP) algorithm [ 16 ]. FBCSP involves frequency band selection using specific filters, integrating frequency domain information, and subsequently extracting spatial features. Canonical classification algorithms include Linear Discriminant Analysis (LDA), Logistic Regression (LR), and Support Vector Machine (SVM). However, traditional ML methods usually separate these two steps and rely on manually crafted features guided by human prior knowledge and experience. In recent years, DL has shown remarkable performance in image recognition, natural language processing (NLP) and speech recognition [ 17 – 19 ], driving the rapid development of artificial intelligence (AI). Against this backdrop, DL techniques are rapidly penetrating into BCI research. DL integrates the two steps of feature extraction and pattern classification, and can automatically extract abstract and high-level features from raw EEG signals. This end-to-end learning approach enhances the system’s performance and generalization ability, enabling it to better adapt to variations in subjects and tasks. Schirrmeister et al. [ 20 ] investigated deep convolutional neural networks (CNNs) with different architectures for decoding MI tasks from raw EEG signals. The study suggests that the CNN-based models achieved at least as good performance as the widely used FBCSP algorithm. Lawhern et al. [ 21 ] designed a compact and lightweight CNN, termed EEGNet, and applied it for four BCI paradigms, i.e., P300, error-related negative (ERN), movement-related cortical potential (MRCP), and sensorimotor rhythm (SMR). Wang et al. [ 22 ] introduced an EEG_GENet network, incorporating a feature-level graph embedding approach and combining it with the EEGNet. Following validation on two BCI datasets suggests that the method outperforms existing approaches compared to data-level graph embedding methods. Wu et al. [ 23 ] proposed a parallel multi-scale filter bank CNN (MSFBCNN) for MI decoding. They trained an end-to-end network to comprehensively extract the spatio-temporal features of EEG signals. Furthermore, they embedded the network structure into the EEGNet to validate the effectiveness of the proposed architecture. Chen et al. [ 24 ] designed a multi-branch network structure called FB-Sinc-SCANet, incorporating a filter bank with Sinc-convolution layers. Additionally, they integrated channel self-attention for feature selection (CSA-FS) from four frequency bands. Recently, attention mechanism has attracted widespread attention and has been utilized in the field of BCI. The Transformer model [ 25 ] aims to address challenges faced by traditional DL models such as recurrent neural networks (RNNs) and long short-term memory (LSTM) networks when dealing with long sequential data. Wairagkar et al. [ 26 , 27 ] showed that time series such as EEG signals have long-term dependencies that can be characterized by long-range temporal correlation. Song et al. [ 28 ] introduced a compact EEG-Conformer network to capture long-term dependencies in MI-EEG signals, encapsulating both local and global features within a unified framework for EEG signal classification. Wang et al. [ 29 ] proposed a novel algorithmic structure combining EEG channel (i.e., spatial) attention with an Swin Transformer. This approach effectively leverages internal connections to narrow the gap between channels and targets, capable of capturing time-frequency-spatial features hidden in EEG data. Wang et al. [ 30 ] proposed an end-to-end DL model named SMANet, which effectively integrates a SincNet, a multibranch spatial-temporal CNN and a channel attention mechanism for MI-EEG decoding. It achieved superior accuracies over six state-of-the-art methods on a four-class MI dataset and two two-class MI datasets. In summary, incorporating temporal and channel attention into convolutional networks has attracted a lot of interests, showing great potential in performance improvement [ 24 , 26 – 33 ]. Although the above-mentioned DL models achieved better performance than traditional ML models, there is still space to promote the classification accuracy for building a robust and reliable BCI system. The performance of a DL model depends largely on its structure and the size of training data. For a given dataset, its size can be augmented to a certain extent using some method and the main challenge of applying DL in an MI-BCI consists in creating a suitable classification model. To address the problem, we propose an end-to-end DL network based on multi-scale CNNs and temporal and channel attention mechanisms (MSTCANet) for decoding MI-EEG data. The EEG data are first fed into four parallel DL branches, each of which includes sequentially connected temporal convolution, spatial convolution, and 2-dimentional average pooling layer, followed by a Transformer-based temporal attention block. Outputs from the four branches are concatenated along channel dimension and then an efficient channel attention network (ECANet) is applied to capture the dependency between feature channels. Finally, a CNN is used to predict the label of a single-trial EEG signal for reducing the computational complexity. The main contributions of this paper are as follows. 1) By efficiently incorporating multiscale spatial-temporal convolutions and the Transformer-based attention mechanism, the proposed network not only extracts comprehensive feature information, but also encapsulates both local and global features within a unified framework for MI-EEG signal classification, enabling the model to learn the global correlation of local temporal features. 2) By concatenating the feature signals from four branches and applying an efficient channel attention network (ECANet), the proposed network is capable of achieving the effective fusion of channel features via automatically adjusting weights to focus more on channels containing key information, significantly enhances the representation capability of feature signals. 3) Based on three public MI-BCI datasets comprising a total of 32 participants and four MI tasks, the performance of the proposed network is analyzed thoroughly in terms of decoding accuracy and standard deviation in inter-session scenarios, and compared with that of five state-of-the art (SOTA) networks. Experimental results validated the superiority of the proposed model. 2. Method In this section, we first provide an overview of the proposed network, then detail each component of the network, and finally delineate its training strategy. 2.1. Overview The proposed DL network is illustrated in Fig 1 , which is based on multi-scale filtering and attention mechanism for extracting spatiotemporal features in EEG signals. The network consists of four modules, i.e., a multi-scale CNN module, a temporal self-attention module, a channel attention module and a classification module. They aim at extracting spatiotemporal features in multiple scales, capturing long-range temporal dependencies in EEG signals, fusing the feature maps from convolutional channels and classifying the fused feature signals into task labels, respectively. Download: PNG larger image TIFF original image Fig 1. Structure of the proposed network. F1 and C denote the number of kernels and the number of EEG channels respectively. https://doi.org/10.1371/journal.pone.0361026.g001 2.2. Multi-Scale CNN module The augmented EEG data are fed into the multi-scale CNN module consisting of four parallel CNN units, each of which comprises a temporal convolutional layer, a spatial convolutional layer and an average pooling layer. The temporal convolutional layers have varying kernel sizes, allowing the extraction of temporal features at multiple different scales. The spatial convolutional layers share the same kernel size for capturing spatial dependencies and topological structures among different EEG electrodes. The average pooling layers are used to reduce the dimensionality of feature signals and computational complexity. The parameters of the multi-scale CNN module are listed in Table 1 . Download: PNG larger image TIFF original image Table 1. Parameters of the multi-scale parallel CNN module. tConv1 ~ 4 denote the four temporal convolutional layers, sConv1 ~ 4 represent the four spatial convolutional layers, and aPool1 ~ 4 denote the four average pooling layers. C and T are the numbers of EEG channels and sampling time points, respectively. https://doi.org/10.1371/journal.pone.0361026.t001 During the temporal convolution process, the convolution kernel of different scales (or filter sizes) extracts temporal feature maps corresponding to different frequency bands. Larger-size kernels capture coarse-grain information, whereas smaller-size kernels can retain fine-grain information [ 34 , 35 ]. Thereby, a single-scale convolutional network may be difficult to extract sufficient information for discriminating MI tasks. A CNN module with multi-scale kernel sizes is an effective strategy to preserve both low- and high-frequency information of EEG signals to improve the performance of MI decoding. Based on this idea, most of previous studies [ 31 , 36 , 37 ] created a multi-scale CNN based on filter bank techniques to extract features from multiple fixed frequency sub-bands such as (1–4 Hz), (4–8 Hz), (8–13 Hz) and (13–30 Hz). However, these frequency bands are prone to change for different subjects because of inter-subject variability and the non-stationary nature of EEG signals, and thus multi-scale CNNs with sub-band filtered input signals is not a good choice. Accordingly, a multi-scale CNN module that uses raw EEG signals as common input is adopted to simultaneously process information from different frequency bands. The input signal is of the dimension (C, T), where C represents the number of EEG channels, and T denotes the length of EEG signals (i.e., the number of sampling points). The kernel sizes of the temporal convolutional layers are determined based on the characteristics of EEG signals and the sampling rate [ 26 ], and are set to 64, 40, 26, and 16 respectively. The maximum and minimum kernels (64 and 16) were determined as approximate fs/4 and fs/16 (fs is the sampling rate), then the remaining two kernels were selected between 64 and 16, and were ultimately determined experimentally to be 40 and 26 respectively. The four kernel sizes correspond approximately to the temporal windows of 250ms, 160ms, 100ms and 64ms respectively, and to the frequencies of 4 Hz, 6 Hz, 10 Hz and 16 Hz respectively. Thereby, the four kernels can be used to extract feature signals above 4 Hz, 6 Hz, 10 Hz and 16 Hz respectively. To maintain the same input and output sizes for subsequent concatenation, padding is set to half the size of the convolutional kernel. Regarding the spatial convolutional layers, the same convolutional kernel with size (C, 1) and stride (1, 1) is utilized to serve as spatial filters. The batch normalization is followed to maintain data stability. The average pooling layers are used to compress the temporal information of the spatially filtered signals and reduce the number of trainable parameters. Their kernel size is set to (1, 75) with a stride of (1, 15). After the multi-scale CNN module, the outputs of the four CNN branches are identical in size, facilitating subsequent feature fusion by concatenating them in channel dimension. 2.3. Temporal self-attention module CNNs may only extract local temporal features, potentially failing to capture long-term dependencies in EEG signals. To address this limitation, we introduce the temporal self-attention module to learn the global temporal correlations of EEG signal features, thereby enlarging the limited receptive field of the single-scale CNN branches. The temporal self-attention module consists of four Transformer blocks, each of which follows a CNN branch. Attention-based Transformer has become a significant milestone in NLP and other sequence modeling tasks [ 25 ]. It has demonstrated outstanding performance in various tasks. Its innovative structure has garnered widespread interest in the scientific community. By appending a Transformer block after a CNN branch, both local and global features are encapsulated within a unified EEG signal classification framework, facilitating learning global correlations in local temporal features. Fig 2(a) shows the architecture of the Transformer block, which has two subblocks: the first one includes a normalization layer and a multi-head attention layer, whereas the second includes a normalization layer and a position-wise fully connected feed-forward layer. The residual connection is used around each of the two subblocks. As shown in Table 1 , the output feature map of a CNN unit has the dimension (40, 1, T1), which is arranged as the feature map with the dimension (T1, 40) by first squeezing the EEG channel dimension and then applying a transpose operation. All convolutional channels at each point in the temporal dimension are treated as a token and are fed into a Transformer block, where the self-attention mechanism is used to learn global temporal dependencies of EEG features, complementing the limited receptive field in a CNN unit. In addition, the multi-head strategy illustrated in Fig 2 (b) is employed in the Transformer block, aiming to further improve representation diversity. Subsequently, the tokens are equally divided into segments (i.e., heads) each with the dimension (T1, 40/h), which are fed into the Transformer block separately. The Transformer begins by transforming the feature map of a single head using a linear layer to obtain Query (Q), Key (K) and Value (V), formulated as (1) where denotes the index of heads, , and are the projection matrices of , and respectively, denotes the divided feature map (i.e., the feature map of i th head), and signifies layer normalization. Download: PNG larger image TIFF original image Fig 2. Transformer block: (a) Structure; (b) Multi-head attention; (c) Scaled dot-product attention. https://doi.org/10
## Related Clinical Research

- [Dynamic VISC Features Improve Japanese Vowel Representation Across Whispered and Normal Phonation](https://medichelpline.com/clinical-feed/plos-one-14-robust-acoustic-representation-of-japanese-vowels-under-phonation-variability.md)
- [Deep Learning 1D-CNN Accurately Classifies Nine Finger Movements from Surface EMG Signals](https://medichelpline.com/clinical-feed/plos-one-8-deep-learning-driven-fine-resolution-finger-movement-recognition-from-surface.md)
- [SPInfer: Fine-tuning Cellpose-SAM for scalable spatial proteomics in human brain tissue](https://medichelpline.com/clinical-feed/biorxiv-1-overcoming-segmentation-bottlenecks-in-brain-tissue-for-spatial-proteomic.md)
- [Cisplatin Modulates Honokiol Effects on Genes Governing Outer Hair Cell Function](https://medichelpline.com/clinical-feed/biorxiv-5-cisplatin-exposure-alters-the-effects-of-honokiol-on-genes-involved-in-outer.md)
- [Understanding the Complexity of Multiple Sclerosis Diagnosis: Lessons from Christina Applegate](https://medichelpline.com/clinical-feed/stat-news-0-opinion-christina-applegate-puts-a-face-on-one-of-the-most-consequential.md)

## Navigation
- [← Back to Neurology Feed](https://medichelpline.com/clinical-feed/neurology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.