Skeletonization of Lagrangian Point Clouds: Extracting Transport Networks from Particle Tracking Data

Viktor Škultéty, Fabrice Guibert, Nicolas Zucker, Thomas Deffieux, Mickael Tanter, Dimitri Van De Ville

Keywords: Point cloud processing; Skeletonization; Flow analysis; Transport networks; Particle tracking

TL;DR: A flow-aware skeletonization method that extracts transport centrelines from particle-tracking data.

Abstract:

Recovering transport pathways from particle-tracking measurements of fluid flows remains challenging due to the large number of detections, directional variability, and the lack of well-defined geometric boundaries. We propose a new algorithm based on a projected $L_1$ geometric median: points are attracted toward local centers of mass while motion along the local mean flow direction is suppressed. This constraint prevents collapse to density maxima without explicit regularization and operates directly on point coordinates, producing skeletons that are not constrained by voxel-grid discretization. Convergence is robust to random initialization given adequate sampling density, and the method remains scalable by operating on an aggregated representation of the data.

We demonstrate feasibility on simulated and experimental data. In particular, for in vivo 3D ultrasound localization microscopy (ULM), the method recovers the microvascular network of a rodent brain. The results indicate that flow-aware skeletonization provides a viable alternative to pipelines operating primarily on particle density representations.

Submission Number: 15