Encode
An implicit-function autoencoder transforms a centerline graph with pseudo-radius information into a compact set of learned tokens.
ECCV 2026 Spotlight
VesselTok
Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation
Figure preview is unavailable.View the original figure ↗
VesselTok method overview · Original figure ↗
Overview
Learning tokens for vessel-like anatomy
Vessel-like structures form intricate spatial networks. VesselTok compresses their geometry into learned latent tokens that support accurate reconstruction and generative modeling without requiring a heavily simplified branch-level representation.
Approach
How VesselTok works
Reconstruct
The learned tokens are decoded into an occupancy field, from which a vessel-like graph can be recovered.
Generate and complete
Generative models operate in the latent space: diffusion supports sampling, and flow matching supports conditional infilling.
Research directions
Applications
Reconstruction
Represent complex vascular and airway geometry using compact learned tokens.
Generation
Learn distributions over three-dimensional vessel-like structures in the latent space.
Completion
Use conditional generative modeling for structural infilling of partial observations.
Resources
Paper, code, and data
The public repository contains implementations of the autoencoder, latent generative models, evaluation, and experiment configurations. Check the repository for installation instructions and model/data availability.
Reference
Citation
@inproceedings{prabhakar2026vesseltok,
title = {VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation},
author = {Prabhakar, Chinmay and Wittmann, Bastian and Amiranashvili, Tamaz and B{\"u}schl, Paul and de la Rosa, Ezequiel and McGinnis, Julian and Wiestler, Benedikt and Menze, Bjoern and Shit, Suprosanna},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026},
note = {arXiv:2603.18797}
}Please check the final proceedings citation information when available.