FM-fMRI generates task brain activity from resting scans, boosting autism detection

Task-based fMRI directly measures neural responses to stimuli but is expensive and hard to collect at scale. To address this, researchers from Yale and collaborators introduce FM-fMRI, a novel generative model that learns a continuous-time conditional vector field to map a subject's resting-state fMRI (rsfMRI) and task event information into realistic task-evoked ROI time series. The use of flow matching enables fast ordinary differential equation (ODE) sampling and flexible conditioning over variable event schedules—key advantages over prior approaches like conditional diffusion, GANs, and VAEs. The model is evaluated not just on pointwise reconstruction but on complementary criteria capturing temporal and spectral structure, connectivity consistency at both subject and group levels, and distributional alignment.

Validated on the public Human Connectome Project and an internal BioPoint autism cohort, FM-fMRI achieves the strongest spectral and connectivity agreement across all baselines. The researchers further demonstrate practical utility by augmenting the limited BioPoint dataset with synthetic task-fMRI signals, which significantly improves downstream autism classification accuracy. This work, early accepted at MICCAI 2026, highlights how generative AI can reduce the acquisition burden of task-fMRI while enabling more robust clinical machine learning models for neurological and psychiatric conditions.