Reproduce “ModeT: Learning Deformable Image Registration via Motion Decomposition Transformer” 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

Haiqiao Wang, Dong Ni, Yi Wang

ModeT registers brain MRI scans by predicting deformable voxel movements. At coarse scales, it uses local attention to produce multiple motion proposals, then combines them through a coarse-to-fine pyramid. On Mindboggle, it reports a Dice score of 62.8% versus 61.1% for PR++, and the highest overlap on both brain-MRI datasets, but does not establish clinical benefit or broader generalization.

Key insights

  • Treating attention weights as correspondence distributions makes it possible to derive candidate displacement fields directly, rather than using attention only for feature extraction.
  • Multiple motion candidates can represent ambiguity at coarse scales, while competitive fusion selects and combines them using local context.
  • The evaluation supports the method on two brain MRI datasets, but does not establish generality to other organs or modalities or quantify computational savings.

Summary by alphaXiv.

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