Reproduce “CurvSegFlow: Time-Conditioned Flow Matching for Robust Segmentation of Curvilinear Structures in Noisy Biomedical Images” on OpenResearch

About this paper

Sidi Mohamed Sid'El Moctar, Achraf Ait Laydi, Alexandre Beber, Marcus Braun, Zdenek Lansky, Yousef El Mourabit, Helene Bouvrais

CurvSegFlow reframes curvilinear biomedical image segmentation as a time-conditioned, iterative reconstruction process rather than a single-pass mask prediction. Combining flow matching with attention-gated U-Net features and a joint velocity–segmentation loss, it preserves thin structures and connectivity across noisy microtubule, vessel, nerve, and artery datasets, with as few as three Euler inference steps.

Key insights

• Iterative refinement improves structural recovery without requiring the many sampling steps commonly associated with diffusion-based segmentation; additional Euler steps beyond three do not consistently improve accuracy.
• The combined flow, Dice, and weighted BCE objective is important: flow loss alone nearly collapses to background predictions, whereas the full objective addresses both velocity learning and severe class imbalance.
• Attention gates mainly improve precision and centerline preservation, supporting the model’s ability to suppress background responses while retaining thin connected structures.

Summary by alphaXiv.