Reproduce “Diffeomorphic Image Registration with Neural Velocity Field” 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

Kun Han, Shanlin sun, Xiangyi Yan, Chenyu You, Hao Tang, Junayed Naushad, Haoyu Ma, Deying Kong and 1 more

Researchers at the University of California, Irvine, developed Diffeomorphic Image Registration with Neural Velocity Field (DNVF), employing an implicit neural representation and sinusoidal activations to model continuous velocity fields for precise, topology-preserving transformations. A cascaded version, Cas-DNVF, combines a learning-based initial deformation with DNVF for refinement, achieving state-of-the-art accuracy on 3D brain MR datasets, including 0.815 DSC on OASIS and 0.612 DSC on Mindboggle, with low rates of negative Jacobian determinants.

Key insights

  • Periodic sinusoidal activation functions are essential for the neural velocity field to effectively model high-frequency content and intricate local deformations, significantly outperforming ReLU-based alternatives.
  • An MLP-based implicit neural representation of a continuous velocity field, when integrated via Scaling and Squaring, can generate highly accurate and topology-preserving diffeomorphic transformations.
  • Combining a learning-based initial deformation with DNVF's instance-specific optimization for residual deformation offers a robust cascaded framework that balances computational efficiency and precision.

Summary by alphaXiv.

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