Research map / Image generation / Fast diffusion sampling
VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis
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 VasTSD, a 3D diffusion model with a dynamic minimum-spanning-tree state-space (Mamba-style) denoiser and a pre-trained vision embedder that synthesizes angiography volumes from non-angiographic scans.
Problem
Synthesizing angiography from non-contrast 3D scans to avoid contrast agents and radiation, while keeping vessels continuous across slices, which 2D slice-based image translation methods fail to do.
Approach
Builds an input-adaptive minimum spanning tree over feature similarities (Kruskal) to define tree-based state-space scanning for 3D denoising, plus a contrastively pre-trained vision embedder supplying vessel masks and tokens as conditions and cross-slice attention for state transitions.
Key results
- VasTSD reaches 29.86 PSNR / 0.9478 SSIM on ITK-TubeTK T1-Flash to MRA, outperforming SynDiff (28.98 / 0.9446) and other baselines.
- On ISICDM 2020 CT-to-CTA, VasTSD achieves 29.72 PSNR / 0.9384 SSIM... with 27.36 PSNR on the CTA column, the best reported PSNR there.
- 3D connectivity score of 9.59 versus 8.26 for the next best (SynDiff) in the 3D volume evaluation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VasTSD instead of DiffMa | Angiography synthesis · ITKTubeTK T1-Flash | PSNR: 29.12 -> 29.86 | Helped |
| use VasTSD instead of DiffMa | Angiography synthesis · ITKTubeTK T2 | PSNR: 29.36 -> 29.72 | Helped |
| use VasTSD instead of SynDiff | Angiography synthesis · ITKTubeTK T2 | SSIM: 0.9482 -> 0.9384 | Hurt |
| use VasTSD instead of SynDiff | Angiography synthesis · ITKTubeTK T1-MPRAGE | SSIM: 0.9516 -> 0.9454 | Hurt |
| use VasTSD instead of DiffMa | Angiography synthesis · ISICDM 2020 CT | PSNR: 28.94 -> 29.48 | Helped |
| use VasTSD instead of cGAN | Angiography synthesis · ISICDM 2020 CT | SSIM: 0.9163 -> 0.9562 | Helped |
| remove pre-training from the vision embedder (ablation) | VasTSD · ITKTubeTK T1-Flash | PSNR: 29.86 -> 26.36 | Hurt |
| remove pre-training from the vision embedder (ablation) | VasTSD · ISICDM 2020 | SSIM: 0.9562 -> 0.8993 | Hurt |
| replace the vision embedder with BioMedCLIP (ablation) | VasTSD · ITKTubeTK T1-Flash | SSIM: 0.9478 -> 0.9286 | Hurt |
| replace tree scan with Zigzag-scan-B (ablation) | VasTSD · ITKTubeTK T1-Flash | SSIM: 0.9478 -> 0.8216 | Hurt |
| replace tree scan with Local-scan-S (ablation) | VasTSD · ITKTubeTK T2 | PSNR: 29.72 -> 28.54 | Hurt |
Surprising findings
- DiffMa has very low SSIM (0.6936 on ITK-TubeTK T1-Flash) despite competitive PSNR, suggesting metric inconsistency.
Limitations
- Anisotropic medical data resolution and slice orientation effects are not resolved.
- Comparison for 3D synthesis is built by stacking 2D baseline outputs since no prior 3D angiography synthesis method exists.
Experimental setup
Datasets: ITK-TubeTK Topcow2024 ISICDM 2020
Benchmarks: ITK-TubeTK ISICDM 2020 Topcow 2024
Metrics: PSNR SSIM Dice Jaccard Connectivity score
Baselines: SAGAN cGAN SynDiff Fast-DDPM DiffMa BioMedCLIP Zigzag-scan Local-scan
Builds on: Mamba SynDiff S6 InfoNCE
Closest papers by meaning
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.87)
- DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.82)
- MuTri: Multi-view Tri-alignment for OCT to OCTA 3D Image Translation (CVPR 2025 · similarity 0.82)
- Sculpt3D: Multi-View Consistent Text-to-3D Generation with Sparse 3D Prior (CVPR 2024 · similarity 0.82)
- Trade-Offs in Fine-Tuned Diffusion Models between Accuracy and Interpretability (AAAI 2024 · similarity 0.82)
- GaraMoSt: Parallel Multi-Granularity Motion and Structural Modeling for Efficient Multi-Frame Interpolation in DSA Images (AAAI 2025 · similarity 0.82)
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.