A deep learning framework integrating multiscale source-space attention and spatiotemporal feature fusion for magnetic source imaging
Biomedical Signal Processing and Control · 2027-03 · 131 · 111831
DOI: 10.1016/j.bspc.2026.111831 ↗Kurz gefasst
Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2027-03. Verfasst von Yanling Yang und 5 weiteren Autor:innen. Die Arbeit nennt 65 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0). Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Magnetic source imaging (MSI) is a noninvasive neuroimaging technique for reconstructing brain activity from scalp magnetoencephalography (MEG) recordings in vivo. However, due to the limited resolution of MEG from measured sensors, MSI still involves the solution of a highly ill-posed inverse problem. Superior to the traditional localization methods, the deep learning (DL) approaches have been used in the MSI for more accurately mapping the relationships between the sensors and sources. This study aims to develop a DL framework that enhances both spatial and temporal resolution for improved source localization, particularly in complex scenarios involving multiple sources or low signal-to-noise ratios (SNR). In this study, a DL-based MSI localization, multiscale source-space attention-based encoder-decoder network (MSA-EDNet), was proposed to accomplish the accurate localization of brain activities. By implementing the encoding–decoding structure, the spatiotemporal features are extracted by multi-scale and dilated convolutions and the source space dimension is enhanced by the source-space attention mechanism (SSAM) during encoding; the deconvolution with bidirectional long short-term memory (BiLSTM) is employed to recovery the source features for decoding. For performance evaluation, our proposed MSA-EDNet was compared with eight novel methods of MNE, dSPM, sLORETA, eLORETA, LCMV, ConvDip, LSTM, and SSINet on simulated and real MEG datasets, respectively. It was demonstrated that our approach outperformed other MSI algorithms, and achieved the lowest localization error (LE), the highest area under the curve (AUC) of (5.305 ± 4.953 mm), (0.961 ± 0.027) for the single source, and the highest AUCs of (0.947 ± 0.093), (0.901 ± 0.099) for two sources and three sources at the same condition (LNs = 1, SNR = 30 dB) on simulated dataset, respectively. Our proposed approach demonstrates robust capacities for betterment localization of MEG brain activities and shows great potentials for the related fields of brain research.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Yang, Y., Hao, Z., Zhao, H., Zhang, Q., Zhou, L., & Yao, X. (2027). A deep learning framework integrating multiscale source-space attention and spatiotemporal feature fusion for magnetic source imaging. Biomedical Signal Processing and Control, 131, 111831. https://doi.org/10.1016/j.bspc.2026.111831
Vancouver
Yang Y, Hao Z, Zhao H, Zhang Q, Zhou L, Yao X. A deep learning framework integrating multiscale source-space attention and spatiotemporal feature fusion for magnetic source imaging. Biomedical Signal Processing and Control. 2027;131:111831. doi:10.1016/j.bspc.2026.111831
BibTeX
@article{Yang2027,
author = {Yang, Yanling and Hao, Zezhou and Zhao, Helong and Zhang, Qingyi and Zhou, Liang and Yao, Xufeng},
title = {A deep learning framework integrating multiscale source-space attention and spatiotemporal feature fusion for magnetic source imaging},
journal = {Biomedical Signal Processing and Control},
year = {2027},
volume = {131},
pages = {111831},
doi = {10.1016/j.bspc.2026.111831}
}
EUVIMED