Reproduce “BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics” 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
Junfeng Xia, Wenhao Ye, Junxiang Zhang, Xuanye Pan, Mo Wang, Quanying Liu
BrainWorld predicts future whole-brain fMRI sequences from recent activity and a person’s structural MRI, conditioning latent-space diffusion on anatomy. Across 22 datasets, it evaluates forecasts up to 400 frames. On external resting-state SALD, it reports MSE 0.218 ± 0.082 and ROI-wise temporal correlation 0.596 ± 0.155, but HCP Movie correlation is 0.081 ± 0.037, limiting claims of fidelity.
Key insights
- Conditioning fMRI generation on structural anatomical priors and framing it as a spatiotemporal problem allows for stable and coherent long-horizon prediction of whole-brain dynamics.
- Injecting sMRI directly into the generative (denoising) process of a DiT, rather than merely concatenating features, provides a more effective multimodal fusion mechanism for fMRI generation and representation learning.
- Leveraging Diffusion Transformers, capable of handling high-dimensional spatiotemporal data and flexible condition injection, is well-suited for modeling complex 4D fMRI dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.