NeuroCycle: Physiologically Constrained Cycling for Generating Neural Information-Rich fMRI from EEG

Yufan Hu, Yujiang, Shengyuan Liu, Yiyang LIN, Cheng Wang, Yixuan Yuan

Keywords: EEG; FMRI; Cross-modal generation; Brain signal synthesis

Abstract: Electroencephalography (EEG) provides millisecond-level temporal resolution but suffers from poor spatial precision, whereas functional magnetic resonance imaging (fMRI) offers fine-grained spatial detail at the expense of cost and latency. Leveraging their complementarity, an emerging direction is to synthesize fMRI from EEG, enriching EEG with spatial information while retaining its efficiency. However, existing EEG to fMRI generation methods often lack designs to preserve information completeness and neglect neurophysiological priors, leading to reconstructions that may appear plausible but fail to ensure neuroscientific validity. We introduce NeuroCycle, a cyclic EEG–fMRI generation framework that enforces information completeness and neuroscientific plausibility. It incorporates two neurophysiological priors: (i) a Cross-Modal ROI-wise Structural Module that aligns fMRI embeddings with EEG-derived correlation patterns to preserve regional organization, and (ii) an R2E Physiological Connectivity Guidance Module that supervises covariance matrices via Riemannian-to-Euclidean mapping to maintain functional connectivity. The bidirectional cycle (EEG$\rightarrow$fMRI$\rightarrow$EEG) further enforces information completeness and cross-modal alignment, ensuring that synthesized fMRI retains key neural information. Experiments on NODDI and Oddball datasets show consistent improvements over state-of-the-art baselines, producing sharper voxel-wise fMRI with richer neural information, preserved connectivity, and stronger cross-modal alignment.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 10230