BrainWorld generates 4D brain activity from structural MRI scans

BrainWorld, developed by Junfeng Xia and team, is a structural-prior-conditioned generative model for whole-brain 4D fMRI dynamics. Unlike existing fMRI foundation models that focus on representation learning or prediction, BrainWorld uses an individual's structural MRI as anatomical context to guide the generation of future brain activity. The structural prior is integrated directly into the denoising process of a diffusion-based framework, making it a condition-aware generator rather than a parallel modality. Evaluated on 22 diverse datasets covering different cohorts and brain states, BrainWorld generates stable fMRI trajectories of up to 400 time frames, significantly longer than typical generative models.

Beyond generation, BrainWorld demonstrates practical utility: using its synthetic samples for data augmentation improves performance on downstream tasks like neurological disorder classification and cognitive state decoding. The learned multimodal representations also transfer better across tasks than baselines. This work positions BrainWorld as a versatile framework for long-horizon brain dynamics modeling and multimodal learning, with potential applications in computational neuroscience, clinical diagnostics, and brain-computer interfaces.