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

Area of Science:

• Cognitive Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Generative adversarial networks (GANs) are used for data augmentation, but simulating functional magnetic resonance imaging (fMRI) data, especially for mild cognitive impairment (MCI), is challenging.
• Characterizing brain function variations in cognitive decline is critical for understanding neurological disorders.

Purpose of the Study:

• 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).
• To develop an innovative method for generating realistic fMRI data that captures temporal dynamics and aids in cognitive neuroscience research.

Main Methods:

• Propose WGTMM, a generative adversarial network (GAN) integrating the Vision Transformer for fMRI (VTFF).
• WGTMM directly generates fMRI time-series data from pink noise, preserving temporal dynamics.
• Utilize a Wasserstein GAN (WGAN) with feature matching to improve generation quality and prevent mode collapse.

Main Results:

• WGTMM-generated fMRI data show lower Kullback-Leibler (KL) divergence than traditional GAN and WGAN models, closely resembling real data.
• Synthetic data generated by WGTMM significantly improve multi-class classification performance when used for data augmentation.
• Analysis of VTFF class token attention patterns revealed monotonic weight variations in key cortical areas across disease stages.

Conclusions:

• WGTMM effectively enriches training datasets for cognitive decline research.
• The method offers new insights into spatial biomarkers of cognitive decline.
• Fine-grained exploration of disease progression is achieved by identifying specific cortical areas with altered activity patterns across cognitive stages.