Reproduce “Omni-fMRI: A Universal Atlas-Free fMRI Foundation Model” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mo Wang, Wenhao Ye, Junfeng Xia, Junxiang Zhang, Xuanye Pan, Minghao Xu, Haotian Deng, Hongkai Wen and 1 more

Omni-fMRI is an atlas-free fMRI foundation model that directly processes voxel-level signals via a dynamic patching mechanism, establishing state-of-the-art performance across diverse neuroimaging tasks with enhanced computational efficiency and interpretable representations.

Key insights

  • Dynamic patching using intensity variance effectively addresses the high dimensionality and non-uniform information density of fMRI, enabling scalable voxel-level processing with global attention.
  • Preserving fine-grained voxel-level information and capturing long-range dependencies through an atlas-free approach leads to more generalizable and robust brain representations across tasks and populations.
  • A reconstruction-based self-supervised objective (MAE) is more effective for fMRI data than joint-embedding architectures, and omitting per-patch normalization can surprisingly improve performance by maintaining biological semantics.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.