Research map / Video / Video segmentation and tracking
DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation
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 DeNVeR, an unsupervised test-time-trained framework that segments coronary vessels in X-ray angiography videos using Hessian priors, layer separation via MLPs, RAFT optical flow, and Eulerian motion fields, and releases the manually labeled XACV dataset.
Problem
Vessel segmentation in X-ray coronary angiography videos without annotated ground truth, where supervised models generalize poorly across imaging systems and frame-by-frame methods lack temporal coherence.
Approach
Test-time training on each test video that decomposes frames into a static background (B-spline heartbeat motion, MLP canonical image) and a foreground vessel layer (CNN canonical image with Eulerian motion), guided by a Hessian-filter prior, a parallel vessel motion loss aligning flow with vessel direction, and flow-warp and mask-consistency losses.
Key results
- On XACV, DeNVeR reaches Jaccard 0.584 and Dice 0.733, versus 0.506/0.660 for the best self-supervised baseline FreeCOS (about 7.7% and 7.3% gains).
- DeNVeR improves over the traditional Hessian filter by 16.9% Jaccard and 14.9% Dice, while remaining unsupervised.
- Ablation: removing the layer separation bootstrap, Hessian prior loss, or parallel loss lowers Dice by 0.0877, 0.1006, and 0.0368 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeNVeR instead of FreeCOS | XACV vessel segmentation · XACV test set | Jaccard: +7.7% | Helped |
| Use DeNVeR instead of FreeCOS | XACV vessel segmentation · XACV test set | Dice: +7.3% | Helped |
| Use DeNVeR instead of Hessian | XACV vessel segmentation · XACV test set | Jaccard: 16.9% improvement | Helped |
| Use DeNVeR instead of Hessian | XACV vessel segmentation · XACV test set | Dice: 14.9% increase | Helped |
| Add layer separation bootstrapping (ablation) | DeNVeR · XACV test set | Dice: 0.0877 | Helped |
| Add Hessian prior loss (ablation) | DeNVeR · XACV test set | Dice: 0.1006 | Helped |
| Add parallel vessel motion loss (ablation) | DeNVeR · XACV test set | Dice: 0.0368 | Helped |
Surprising findings
- DeNVeR's unsupervised test-time method performs on par with the supervised U-Net in some sequences, which the authors attribute to U-Net overfitting on limited data.
- The self-supervised SSVS scored slightly below the plain Hessian filter under the chosen thresholds.
Limitations
- Sensitive to preprocessing Hessian filter, which may misidentify non-vessel structures.
- Substantial computational overhead of about 20 minutes per 80-frame video.
- Restricted to contrast-enhanced vessel videos; not applicable to static vessel images such as retinal scans.
Experimental setup
Datasets: XACV CADICA
Benchmarks: XACV CADICA
Metrics: Jaccard Dice accuracy sensitivity specificity clDice NSD AUROC AUPRC
Baselines: Hessian filter U-Net SSVS DARL FreeCOS
Builds on: RAFT NIR U-Net Hessian-based vesselness filter Eulerian motion
Closest papers by meaning
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.87)
- DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image (CVPR 2025 · similarity 0.85)
- Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel Segmentation (CVPR 2025 · similarity 0.85)
- VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.82)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.82)
- Label-Efficient Data Augmentation with Video Diffusion Models for Guidewire Segmentation in Cardiac Fluoroscopy (AAAI 2025 · 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.