Research map / 3D vision / Depth and stereo
DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeformCL, a continuous graph-based deformable centerline representation that is iteratively deformed from an adaptive template and used to produce distance-map-based 3D vessel segmentations, plugged into softDice, clDice and DSCNet.
Problem
Discrete per-voxel mask representations for 3D vessel extraction cause fragmented segments, poor noise robustness, and weak feature aggregation along tubular structures.
Approach
Represent vessels as a graph of centerline points that is adaptively initialized from the coarse segmentation and refined through cascaded Transformer-based deformation and unpooling, then convert the centerlines to segmentation via a distance map; trained with local chamfer, SDF, and edge-length regularization losses.
Key results
- On HaN-Seg, Ours (softDice) improves Dice from 76.30 to 79.58 and clDice from 85.96 to 88.86 over the softDice baseline.
- On ImageCAS, Ours (DSCNet) reaches Dice 88.84 and clDice 96.38, up from 87.98 and 95.34 for DSCNet.
- Centerline F1 on HaN-Seg rises from 91.17 (softDice) to 94.56 with DeformCL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DeformCL to softDice | softDice · HaN-Seg | Dice: 76.30 -> 79.58 | Helped |
| Add DeformCL to softDice | softDice · HNCTA | Dice: 83.05 -> 84.69 | Helped |
| Add DeformCL to softDice | softDice · ASOCA | Dice: 84.06 -> 85.16 | Helped |
| Add DeformCL to softDice | softDice · ImageCAS | Dice: 87.59 -> 88.33 | Helped |
| Use adaptive template generation with k=4 rather than a straight initial template (ablation) | DeformCL · HaN-Seg | Dice: 78.87 -> 79.58 | Helped |
| Increase template control points from k=4 to k=8 (ablation) | DeformCL · HaN-Seg | Dice: 79.58 -> 78.36 | Hurt |
| Increase template control points from k=4 to k=16 (ablation) | DeformCL · HaN-Seg | Dice: 79.58 -> 77.22 | Hurt |
| Increase template control points from k=4 to k=64 (ablation) | DeformCL · HaN-Seg | Dice: 79.58 -> 77.37 | Hurt |
| Use local rather than global chamfer distance loss (ablation) | DeformCL · HaN-Seg | clDice: 86.62 -> 88.86 | Helped |
| Add regularization loss (ablation) | DeformCL · HaN-Seg | Dice: 78.38 -> 79.58 | Helped |
| Add SDF loss (ablation) | DeformCL · HaN-Seg | Dice: 79.51 -> 79.58 | No clear effect |
Surprising findings
- Overly accurate initial templates (more control points, k=8 to 64) hurt performance, so a coarse template works better.
- Global chamfer distance loss causes larger training fluctuations than the proposed local chamfer loss.
Limitations
- PointScatter and Topoloss were not compared because they are 2D methods and GPU memory limits made them impractical for 3D.
- Gains on some metrics are small or mixed, e.g. ImageCAS clDice with the clDice base improves by only 0.01 and some Betti errors worsen.
Experimental setup
Datasets: HaN-Seg ASOCA ImageCAS HNCTA
Benchmarks: HaN-Seg ASOCA ImageCAS HNCTA
Metrics: Dice clDice Betti error β0 Betti error β1 Euler characteristic error HD95 F1 Chamfer distance
Baselines: softDice clDice DDT DSCNet
Builds on: 3D UNet Transformer clDice DSCNet Pixel2Mesh
Closest papers by meaning
- DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.85)
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.85)
- Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel Segmentation (CVPR 2025 · similarity 0.82)
- GraphMorph: Tubular Structure Extraction by Morphing Predicted Graphs (NeurIPS 2024 · similarity 0.82)
- VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.81)
- H-ViT: A Hierarchical Vision Transformer for Deformable Image Registration (CVPR 2024 · similarity 0.81)
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.