FlexiBrain: New AI framework processes raw fMRI data 10x faster
Outperforms state-of-the-art by 12 points without any data augmentation—
FlexiBrain, a new framework from Mo Wang and colleagues, tackles a core challenge in fMRI analysis: severe data heterogeneity. Traditional methods require rigid preprocessing pipelines that warp native brain scans into a uniform grid, risking loss of subject-specific anatomical info and burning hours per subject. FlexiBrain instead defines patch sizes in real-world physical units and applies dynamic patch resizing, allowing it to ingest data in its native resolution without spatial standardization.
Powered by a Mamba-JEPA backbone designed for high-dimensional 4D fMRI signals, FlexiBrain consistently outperforms recent state-of-the-art methods across five diverse neuroscience tasks—by up to 12 percentage points—without any external data augmentation. It also acts as a seamless plug-in module, drastically reducing preprocessing overhead. For researchers building robust voxel-level fMRI foundation models, this means faster iteration and more anatomically faithful representations.
- FlexiBrain uses dynamic patch resizing in real-world physical units, eliminating destructive spatial standardization for native fMRI data.
- Mamba-JEPA backbone efficiently models high-dimensional 4D fMRI signals, achieving up to 12 percentage point gains over SOTA across five tasks.
- Acts as a plug-in module that reduces preprocessing from hours to minutes, accelerating development of robust fMRI foundation models.
Why It Matters
Faster, more accurate fMRI analysis that preserves individual brain anatomy, enabling scalable neuroscience research and clinical applications.