Binary Flow Matching for Vascular Segmentation
A vascular segmentation method that combines Binary Flow Matching with Straight-Path Regularization to improve connectivity and inference efficiency.
Overview
This work proposes Binary Flow Matching (BFM) to address disconnected and ambiguous thin-vessel segmentation in medical images.
It combines a generation process tailored to binary masks with efficient path learning to improve vascular connectivity and inference efficiency.
Problem
- Thin vessels and complex branches are easily disconnected during segmentation
- Vascular structures must be recovered reliably around ambiguous boundaries
- Generative segmentation models need to balance accuracy and inference efficiency
Approach
- A Binary Flow Matching process designed for binary-mask generation
- Straight-Path Regularization for learning more direct generation trajectories
- Evaluation of segmentation quality and structural connectivity on three vascular datasets
Results
- Accepted to MICCAI 2026 as a first-author paper
- Outperformed comparison methods across three vascular datasets
- Improved vascular connectivity and generative inference efficiency