Reproduce “Flow Matching Meets 3D Curvilinear Structure Segmentation in Medical Imaging” 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
Sidi Mohamed Sid'El Moctar, Nicolas Vitry, Hélène Bouvrais
3D-CurvSegFlow reformulates vascular segmentation as deterministic flow matching, progressively transforming noise into a vessel mask with a time-conditioned 3D U-Net. Using only three Euler integration steps, the same 23.49-million-parameter model achieves Dice scores of 0.74 on portal-vein CT, 0.8032 on cerebral TOF-MRA, and 0.824 on coronary CTA, with particularly strong recall and centerline preservation across the three datasets.
Key insights
- Iterative refinement improves more than volumetric overlap: the method strengthens vessel recall and centerline continuity, helping recover fine branches and reduce fragmented vascular trees.
- The same general architecture performs across abdominal CT, high-resolution cerebral MRA, and coronary CTA despite differences in anatomy, resolution, artifacts, and vessel scale.
- More integration steps do not automatically improve predictions; three steps performed better than five or ten in the reported experiment, suggesting that the model is coupled to its training-time discretization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.