Research map / Graph learning / Graph self-supervised learning
VesselTok: Tokenizing Vessel-like 3D Biomedical Graph Representations for Reconstruction and Generation
Chinmay Prabhakar, Bastian Wittmann, Tamaz Amiranashvili, Paul Büschl, Ezequiel De la Rosa, Julian McGinnis, Benedikt Wiestler, Bjoern Menze, Suprosanna Shit
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| tokenize vessel centerline occupancy instead of surface points (VesselTok vs 3DShape2VecSet/Hunyuan3D) | VAE tokenizer · six anatomies | clDice / CD / |Δβ0|: e.g. COSTA clDice 59.65 -> 77.26; ATM 90.08 -> 96.33 | Improved |
| VesselTok vs 3DShape2VecSet on loop error | VesselTok · HiPas / PARSE / Pulmonary-AV | |Δβ1|: HiPas 28.65 -> 38.42; PARSE 32.31 -> 34.06; Pulm 5.04 -> 5.52 | Hurt |
| VesselTok vs vessel-specific VesselGPT | VesselTok · ATM | clDice: 77.09 -> 96.61 | Improved |
| VesselTok vs VesselGPT on connected components | VesselTok · ATM | |Δβ0|: 0.05 -> 0.07 | Hurt |
| apply VesselTok to unseen anatomies and scales | VesselTok · TopCoW / renal / synthetic cuts | clDice: TopCoW 97.96 -> 99.42; RV 79.66 -> 88.86 | Improved |
| use VesselTok tokens for latent diffusion | EDM · generation | FID: cond 73.58 -> 43.13; uncond 146.57 -> 96.87 | Improved |
| use VesselTok latent for infill vs autodecoder | VesselTok · ATM infill | clDice: 83.49 -> 88.13 | Improved |
| increase number of tokens K from 512 to 768 (ablation) | VesselTok · ATM | clDice: 96.61 -> 97.13; compression 7.03 -> 4.68 | No clear effect |
| reduce tokens K from 512 to 256 or fewer (ablation) | VesselTok · ATM | clDice: 96.61 -> 80.35 (256); 12.29 (128); 8.42 (64) | Hurt |
| increase channel dim C from 4 to 16 (ablation) | VesselTok · ATM | clDice: 96.61 -> 96.32; compression 7.03 -> 1.76 | No clear effect |
| reduce channel dim C from 4 to 2 (ablation) | VesselTok · ATM | clDice: 96.61 -> 7.08 | Hurt |
| increase pseudo-radius r from 0.008 to 0.032 (ablation) | VesselTok · ATM | clDice loss vs beta1 error: clDice loss ~0.28 -> ~0.007; beta1 error ~4 -> ~16 | Mixed |
Closest papers by meaning
- Whole Brain Vessel Graphs: A Dataset and Benchmark for Graph Learning and Neuroscience (NeurIPS 2021 · similarity 0.89)
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.87)
- GraphMorph: Tubular Structure Extraction by Morphing Predicted Graphs (NeurIPS 2024 · similarity 0.85)
- VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.85)
- DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image (CVPR 2025 · similarity 0.84)
- Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
