Research on fMRI Image Generation from EEG Signals Based on Diffusion Models

Amidrapid advances in intelligent medicine, decoding brain activity from electroencephalogram (EEG) signals has emerged as a critical technical frontier for brain–computer interfaces and medical AI systems. Given the inherent spatial resolution limitations of an EEG, researchers frequently integrate functional magnetic resonance imaging (fMRI) to enhance neural activity representation. However, fMRI acquisition is inherently complex. Consequently, efforts increasingly focus on cross-modal transformation methods that map EEG signals to fMRI data, thereby extending EEG applications in neural mechanism studies. The central challenge remains generating high-fidelity fMRI images from EEG signals. To address this, we propose a diffusion model-based framework for cross-modal EEG-to-fMRI generation. To address pronounced noise contamination in electroencephalographic (EEG) signals acquired via simultaneous recording systems and temporal misalignments between EEGs and functional magnetic resonance imaging (fMRI), we first apply Fourier transforms to EEG signals and perform dimensionality expansion. This constructs a spatiotemporally aligned EEG–fMRI paired dataset. Building on this foundation, we design an EEG encoder integrating a multi-layer recursive spectral attention mechanism with a residual architecture.In response to the limited dynamic mapping capabilities and suboptimal image quality prevalent in existing cross-modal generation research, we propose a diffusion-model-driven EEG-to-fMRI generation algorithm. This framework unifies the EEG feature encoder and a cross-modal interaction module within an end-to-end denoising U-Net architecture. By leveraging the diffusion process, EEG-derived features serve as conditional priors to guide fMRI reconstruction, enabling high-fidelity cross-modal image generation. Empirical evaluations on the resting-state NODDI dataset and the task-based XP-2 dataset demonstrate that our EEG encoder significantly enhances cross-modal representational congruence, providing robust semantic features for fMRI synthesis. Furthermore, the proposed cross-modal generative model achieves marked improvements in structural similarity, the root mean square error, and the peak signal-to-noise ratio in generated fMRI images, effectively resolving the nonlinear mapping challenge inherent in EEG–fMRI data.