Integration of multi-scale encoder and convolutional Transformer for epileptic classification(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第12期
- Page:
- 1621-1631
- Research Field:
- 医学信号处理与医学仪器
- Publishing date:
Info
- Title:
- Integration of multi-scale encoder and convolutional Transformer for epileptic classification
- Author(s):
- ZHAI Fengwen1; GAO Kangxia1; JIN Jing1; WEI Bangcai2
- 1. School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China 2. China Telecom Wanwei Information Technology Co., Ltd., Lanzhou 730000, China
- Keywords:
- Keywords: electroencephalogram signal epilepsy deep learning MTSDCformer information purification unit
- PACS:
- R318；TP391
- DOI:
- DOI:10.3969/j.issn.1005-202X.2025.12.012
- Abstract:
- To address the limitation that most Transformer-based models fail to fully learn the rich spatiotemporal information in electroencephalogram (EEG) signals, this study designs a multi-scale spatiotemporal convolutional encoder module and a dense convolutional Transformer (DCT) module, and integrates these two modules to propose MTSDCformer which enables the prediction and classification of EEG signals in epilepsy patients. The model uses a multi-scale spatiotemporal feature encoder to extract feature information from the segmented EEG signals across both the temporal and channel dimensions, and integrates convolutional branches dedicated to local information learning into the Transformer encoder, thereby forming the DCT module. Through the utilization of multi-layers of the DCT module, the model captures the global dependencies and finer-grained temporal features in EEG signals. Furthermore, an information purification unit is employed to encode the frequency information inherent in the EEG signals, facilitating the extraction of detailed features from the intermediate layers. Finally, based on the extracted features, the model classifies the input EEG signals to determine whether the patient has epilepsy. The effectiveness of the model is evaluated on the UPenn & Mayo Clinic Challenge dataset and the public Childrens Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) dataset. Experimental results demonstrate that the MTSDCformer model outperforms baseline methods including LSTM, Vision Transformer, Compact Convolution Transformer, EEG-Conformer, EEGNet, R-Transformer, and ARNN.
Last Update: 2025-12-29