MS-DBNet

MS-DBNet: A Heterogeneous Temporal Convolutional Network for Robust Subject-Specific Cross-Session Motor Imagery Decoding

This is the official code repository for the paper:

Fenglin Shi "MS-DBNet: a heterogeneous temporal convolutional network for robust subject-specific cross-session motor imagery decoding", Proc. SPIE 13940, International Conference on Machine Learning, Neural Networks, and Computer Software (MLNNCS 2025), 139400M (19 November 2025); https://doi.org/10.1117/12.3092280

---

1. Introduction

The practical application of motor imagery brain-computer interfaces (MI-BCIs) is hindered by the challenge of cross-session generalization. To address this, we propose MS-DBNet — a heterogeneous temporal convolutional network that employs a parallel dual-branch architecture to complementarily fuse fine-grained local temporal features with robust global contextual features, thereby enhancing robustness against inter-session signal drift.

Cross-session evaluations on five public datasets demonstrate that MS-DBNet significantly outperforms existing baseline models, achieving an average accuracy of 86.03% on BCIC IV 2a.

2. Model Architecture

Figure 1: Overall architecture of MS-DBNet.

MS-DBNet consists of two parallel branches:

Branch	Name	Role
SSTB	Single-Scale Temporal Branch	Employs spatial depthwise convolution and single-scale temporal depthwise separable convolution to precisely capture fine-grained local temporal patterns at high resolution.
MSDB	Multi-Scale Dilated Branch	Utilizes multi-scale kernels and dilated convolutions to systematically expand the receptive field, learning long-range contextual dependencies tolerant to temporal variations.

Both branches have the same front-end (Block 1 temporal convolution + Block 2 spatial depthwise convolution), but diverge in their temporal processing blocks:

	SSTB	MSDB
Block 1	Conv2d, F1=16, k=(1,32)	Conv2d, F'1=16, k=(1,32)
Block 2	DepthwiseConv, D=8, k=(C,1), Pool=(1,8)	DepthwiseConv, D'=2, k=(C,1), Pool=(1,8)
Block 3	SeparableConv, F3=32, k=(1,32), Pool=(1,4)	MultiScaleConv, F'3=64, k={3,7,15,31}, Pool=(1,4)
Block 4	Conv2d, F4=32, k=(1,3), Pool=(1,1)	DilatedMultiScaleConv, F'4=64, d={1,2,4,8}, Pool=(1,1)

All convolutional layers are followed by BN and ELU. All core blocks are followed by a hybrid Channel-Time Attention (CTA) and Dropout (p=0.5). After flattening, the features from both branches are concatenated and fed into a fully-connected classifier.

3. Experimental Results

Cross-Session Performance Comparison (Mean % ± Std. Dev.)

Dataset	FBCSP	EEGNet	DeepConvNet	ShallowConvNet	ATCNet	MS-DBNet (ours)
BCIC IV 2a	59.07 ±14.27	66.67 ±7.90	69.52 ±12.37	71.45 ±12.02	76.20 ±8.79	86.03 ±8.01
BCIC IV 2b	67.82 ±13.80	75.56 ±10.87	73.10 ±13.71	68.35 ±13.21	76.52 ±12.64	77.85 ±11.18
OpenBMI	60.55 ±14.03	63.96 ±13.35	65.34 ±13.99	70.44 ±15.00	70.96 ±15.94	74.14 ±14.56
SHU 2C	55.42 ±8.55	80.83 ±13.63	80.33 ±13.40	80.15 ±13.35	78.62 ±14.08	82.79 ±14.63
SHU 3C	39.95 ±7.95	68.42 ±15.28	70.73 ±10.85	67.98 ±12.29	69.91 ±14.67	74.57 ±11.64
Average	56.56 ±10.32	71.09 ±6.93	71.80 ±5.53	71.67 ±4.95	74.44 ±3.79	79.08 ±5.20

Ablation Study on BCIC IV 2a

Model Variant	Accuracy (%)	F1-Score (%)	Kappa
MS-DBNet (full model)	86.03 ±8.01	86.05 ±8.00	0.814 ±0.107
w/o MSDB (SSTB-only)	79.44 ±9.54	79.37 ±9.56	0.726 ±0.127
w/o SSTB (MSDB-only)	80.09 ±6.08	80.00 ±6.14	0.735 ±0.081
w/o Multi-Scale	82.83 ±7.24	82.72 ±7.33	0.771 ±0.097
w/o Dilation	82.52 ±7.74	82.36 ±7.91	0.767 ±0.103
w/o Attention	83.83 ±6.92	83.68 ±7.04	0.785 ±0.092

Figure 2: Accuracy comparison of MS-DBNet, SSTB-only, and MSDB-only for the 9 subjects in the BCIC IV 2a dataset.

4. Repository Structure

MS-DBNet/
├── README.md                     # User Manual
├── ms_dbnet.py                   # Model definitions (MSDBNet, SSTB, MSDB)
├── modules.py                    # Core building blocks (CTA, MultiScaleTemporalConv, etc.)
└── figures/
    ├── MS-DBNet.png              # Architecture diagram
    └── Accuracy comparison.png   # Comparison diagram

File Descriptions

File	Description
ms_dbnet.py	Contains the three model classes: MSDBNet (full dual-branch model), SSTB (Single-Scale Temporal Branch), and MSDB (Multi-Scale Dilated Branch). Each branch can be used independently for ablation studies.
modules.py	Contains the five reusable building blocks: Conv2dWithConstraint, LinearWithConstraint, ChannelTimeAttention (CTA), MultiScaleTemporalConv, and DilatedMultiScaleConv.

5. Citation

If you find this work useful, please cite:

@inproceedings{shi2025msdbnet,
    title     = {MS-DBNet: A heterogeneous temporal convolutional network for
                 robust subject-specific cross-session motor imagery decoding},
    author    = {Shi, Fenglin},
    booktitle = {International Conference on Machine Learning, Neural Networks,
                 and Computer Software (MLNNCS 2025)},
    volume    = {13940},
    pages     = {139400M},
    year      = {2025},
    publisher = {SPIE},
    doi       = {10.1117/12.3092280}
}

6. Contact

For questions or issues, please open a GitHub issue or contact: shifenglin@connect.hku.hk