EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

WGTMM: WGAN with Transformer Feature Matching for Generating fMRI Data in MCI Patients

Bocheng Wang iD

Brain Sciences · 25.06.2026 · 16 (7) · 665

DOI: 10.3390/brainsci16070665 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Brain Sciences, online erschienen am 25.06.2026. Verfasst von Bocheng Wang. Beteiligte Einrichtung: School of Media Engineering, Communication University of Zhejiang, No.998, Xueyuan Street, Hangzhou 310018, China. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: NIA NIH HHS, National Natural Science Foundation of China, China Scholarship Council.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

The emergence of generative adversarial networks has laid the groundwork for data augmentation, addressing challenges of missing training data in various research scenarios. However, simulating functional magnetic resonance imaging (fMRI) data remains particularly challenging, especially for populations with varying degrees of mild cognitive impairment (MCI). Effectively characterizing and capturing the mechanisms of brain function variations poses a critical issue in cognitive neuroscience. This study aims to simulate and analyze synthetic fMRI blood-oxygen-level-dependent (BOLD) signals across four cognitive stages: healthy control (HC), early MCI (EMCI), late MCI (LMCI), and Alzheimer’s disease (AD).

Methods

We propose WGTMM, an innovative method that integrates the Vision Transformer for fMRI (VTFF) into a generative adversarial network architecture. Crucially, WGTMM directly generates fMRI time-series data from pink noise rather than modeling in a latent space, thereby preserving rich temporal dynamics. The framework incorporates a Wasserstein GAN (WGAN) with feature matching to enhance generation quality and mitigate mode collapse. Results: demonstrate that WGTMM-generated fMRI data exhibit lower Kullback-Leibler (KL) divergence compared to traditional GAN and WGAN models, indicating a closer resemblance to real datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Furthermore, when applied to data augmentation, the synthetic data substantially improve multi-class classification performance.

Conclusions

WGTMM not only enriches training datasets but also provides new insights into spatial biomarkers of cognitive decline. By leveraging VTFF to investigate class token attention patterns across 360 brain regions, this study reveals monotonic weight variations along disease stages in key cortical areas, including the rostral Area 6, the primary sensory cortex, and PFm near Wernicke’s area, offering a fine-grained exploration of disease progression.

Schlagwörter

Stichwörter der Autor:innen

Alzheimer’s diseaseGANTransformerViTfeature matching

Autor:innen und Institutionen

  • School of Media Engineering, Communication University of Zhejiang, No.998, Xueyuan Street, Hangzhou 310018, China.

ORCID-Profile: Bocheng Wang

Förderung

  • NIA NIH HHS
  • National Natural Science Foundation of China
  • China Scholarship Council
  • Zhejiang Provincial Natural Science Foundation

Zitieren

APA

Wang, B. (2026). WGTMM: WGAN with Transformer Feature Matching for Generating fMRI Data in MCI Patients. Brain Sciences, 16(7), 665. https://doi.org/10.3390/brainsci16070665

Vancouver

Wang B. WGTMM: WGAN with Transformer Feature Matching for Generating fMRI Data in MCI Patients. Brain Sciences. 2026;16(7):665. doi:10.3390/brainsci16070665

BibTeX
@article{Wang2026,
  author = {Wang, Bocheng},
  title = {WGTMM: WGAN with Transformer Feature Matching for Generating fMRI Data in MCI Patients},
  journal = {Brain Sciences},
  year = {2026},
  volume = {16},
  number = {7},
  pages = {665},
  doi = {10.3390/brainsci16070665}
}
Kostenlos starten →