Temporal Diffusion Bridge Model for Vessel Extraction Using Physics-informed Synthetic Data

Jinkui Hao; Donald R. Cantrell; Ramez Abdalla; Sameer A. Ansari; Bo Zhou

eess.IVOct 8, 2026

Temporal Diffusion Bridge Model for Vessel Extraction Using Physics-informed Synthetic Data

Authors: Jinkui Hao, Donald R. Cantrell, Ramez Abdalla, Sameer A. Ansari, Bo Zhou

Organizations: Department of Radiology, Northwestern University, Chicago, IL, USA

Abstract

Accurate artery extraction from X-ray coronary angiography (XCA) is essential for coronary artery disease diagnosis and intervention. However, low vessel visibility, respiratory and cardiac motion, and limited annotations challenge existing methods. Moreover, most deep learning approaches produce binary vessel masks without preserving clinically relevant intra-arterial contrast information. We propose a Temporal Diffusion Bridge Model (TDBM) that formulates vessel extraction as image translation and learns from physics-informed synthetic data without manual vessel annotations. A physics-based XCA synthesis engine generates diverse, perfectly aligned pairs of contrast-enhanced and contrast-free images. Using these pairs, TDBM learns cross-domain mappings to recover high-fidelity vessel information. Temporal conditioning incorporates adjacent frames to improve consistency, while a vessel-aware contrastive loss exploits background-invariant synthetic pairs to enhance vascular discrimination. We evaluate TDBM on two XCA vessel segmentation datasets and conduct a reader study using two real-patient XCA sequence datasets. Zero-shot segmentation is further assessed on retinal imaging and fluorescence microscopy datasets. Results demonstrate that TDBM consistently outperforms existing methods and generalizes effectively across imaging modalities. Code and trained weights will be available at: https://github.com/Advanced-AI-in-Medicine-and-Physics-Lab/TDBM

Figures & tables

Fig. 1: Framework overview. (A) A physics-based engine synthesizes realistic, aligned training pairs. (B) The Temporal Diffusion Bridge Model (TDBM) performs vessel extraction as a Brownian Bridge process, using an Attention U-Net conditioned on a temporal window for motion consistency. (C) Training is enhanced by a vessel-aware contrastive loss (VCL) for feature discrimination.

Parameter	Value
Input resolution	$256\times 256$
Temporal radius $\Delta$	$1$ (3 frames)
Diffusion timesteps $T$	$1000$
Reconstruction loss	MSE
VCL weight $\alpha$	$0.01$
VCL temperature $\tau$	$0.05$

Table 1: Key hyperparameters used for TDBM training and inference.

Method	Internal					External
Method	Acc.	Sen.	DSC	clDice	clRec.	Acc.	Sen.	DSC	clDice	clRec.
nnUNet †	0.952 $\pm$ 0.048	0.580 $\pm$ 0.142	0.635 $\pm$ 0.128	0.684 $\pm$ 0.106	0.692 $\pm$ 0.035	0.957 $\pm$ 0.062	0.529 $\pm$ 0.152	0.620 $\pm$ 0.201	0.665 $\pm$ 0.168	0.671 $\pm$ 0.028
MoG-RPCA	0.963 $\pm$ 0.058	0.774 $\pm$ 0.134	0.747 $\pm$ 0.140	0.782 $\pm$ 0.115	0.802 $\pm$ 0.028	0.943 $\pm$ 0.017	0.393 $\pm$ 0.152	0.494 $\pm$ 0.144	0.528 $\pm$ 0.121	0.548 $\pm$ 0.025
MCR-RPCA	0.968 $\pm$ 0.013	0.476 $\pm$ 0.228	0.588 $\pm$ 0.197	0.612 $\pm$ 0.164	0.624 $\pm$ 0.042	0.918 $\pm$ 0.025	0.020 $\pm$ 0.032	0.031 $\pm$ 0.046	0.038 $\pm$ 0.042	0.041 $\pm$ 0.018
TRPCA	0.965 $\pm$ 0.046	0.648 $\pm$ 0.189	0.695 $\pm$ 0.176	0.729 $\pm$ 0.143	0.744 $\pm$ 0.031	0.933 $\pm$ 0.024	0.164 $\pm$ 0.082	0.267 $\pm$ 0.119	0.291 $\pm$ 0.098	0.301 $\pm$ 0.022
ETRPCA	0.955 $\pm$ 0.023	0.724 $\pm$ 0.097	0.654 $\pm$ 0.110	0.718 $\pm$ 0.088	0.767 $\pm$ 0.026	0.942 $\pm$ 0.017	0.681 $\pm$ 0.118	0.637 $\pm$ 0.069	0.695 $\pm$ 0.061	0.732 $\pm$ 0.019

Table 2: Vessel segmentation results (mean

\pm

std) on the internal and external datasets. The best performance for each metric is highlighted in bold.

***

indicates that the difference in DSC between the proposed method and the state-of-the-art method is statistically significant at

p<0.001

according to the Wilcoxon signed-rank test.

\dagger

indicates that manual annotations were used during training.

Fig. 2: The visualization results of vessel extraction and segmentation on the internal and external datasets are presented. (A) and (a) show the input images and the corresponding ground truth of segmentation, respectively. (B–E) display the vessel extraction results obtained by ETRPCA, DDSA, DOSTA-Net and our TDBM. The corresponding binary segmentation results for each method are shown in (b–e).

Method	Year	Training paradigm	Labeled training data	Dice/F1
CoroSAM a	2026	Supervised	1,200	0.730
FAVS a	2026	Supervised	1,326	0.705
TRACS a	2026	Annotation-free	0	0.616
DOSTA-Net b	2026	Annotation-free	0	0.282
TDBM (Ours)	2026	Annotation-free	0	0.671

Table 3: Comparison of vessel segmentation performance on the DCA1 dataset. “Labeled training data” denotes the number of real XCA images with manual vessel annotations used for training or fine-tuning.

Fig. 3: Overall ranking distribution of the reader study across internal and external datasets. Three board certified radiologists independently ranked the results of each method (DSA, DOSTA-Net, and Ours) based on vessel extraction quality.

Fig. 4: Visualization results of real-sequence XCA images used for subjective evaluation from the internal (a) and external (b) datasets. Results obtained using DSA, DOSTA-Net, and the proposed TDBM are shown separately along the sequence. Red boxes highlight and provide magnified views of the key regions of interest.

No.	Temporal Cond.	Slice No.	VCL	Annealing Term	Training No.	Acc.	Sen.	DSC	clDice	clRec.
1	✗	1	✓	✓	50k	0.958 $\pm$ 0.012	0.776 $\pm$ 0.151	0.727 $\pm$ 0.094	0.744 $\pm$ 0.091	0.801 $\pm$ 0.126
2	✗	3	✓	✓	50k	0.957 $\pm$ 0.014	0.792 $\pm$ 0.096	0.738 $\pm$ 0.060	0.758 $\pm$ 0.061	0.814 $\pm$ 0.091
3 ⋆	✓	3	✓	✓	50k	0.961 $\pm$ 0.011	0.810 $\pm$ 0.084	0.755 $\pm$ 0.049	0.779 $\pm$ 0.051	0.831 $\pm$ 0.078
4	✓	5	✓	✓	50k	0.954 $\pm$ 0.016	0.795 $\pm$ 0.058	0.747 $\pm$ 0.056	0.768 $\pm$ 0.057	0.819 $\pm$ 0.060
5	✓	7	✓	✓	50k	0.949 $\pm$ 0.014	0.798 $\pm$ 0.050	0.746 $\pm$ 0.048	0.766 $\pm$ 0.050	0.821 $\pm$ 0.055
6	✓	3	✓	✗	50k	0.953 $\pm$ 0.015	0.808 $\pm$ 0.052	0.743 $\pm$ 0.052	0.762 $\pm$ 0.054	0.823 $\pm$ 0.057

Table 4: Performance (mean

\pm

std) of the ablation study on the external vessel segmentation dataset.

\star

indicates the default configuration used throughout the main experiments.

Setting	VCL location	DSC	clDice
No VCL	None	0.736 $\pm$ 0.048	0.748 $\pm$ 0.051
Deep encoder	Deepest encoder block (1/8)	0.743 $\pm$ 0.050	0.758 $\pm$ 0.052
Bottleneck	Bottleneck (1/8)	0.746 $\pm$ 0.051	0.762 $\pm$ 0.053
Decoder	Three decoder stages	0.751 $\pm$ 0.050	0.771 $\pm$ 0.052
All levels	Encoder + bottleneck + decoder	0.755 $\pm$ 0.049	0.779 $\pm$ 0.051

Table 5: Ablation on the location of the vessel-aware contrastive loss (VCL) on the external XCA30 dataset (mean

\pm

std).

Fig. 5: Visualizations of representative challenging cases involving interventional artifacts.

Sampling Steps	Accuracy	Sensitivity	Specificity	DSC
2	0.928 $\pm$ 0.024	0.093 $\pm$ 0.069	0.999 $\pm$ 0.002	0.159 $\pm$ 0.106
10	0.961 $\pm$ 0.011	0.799 $\pm$ 0.096	0.973 $\pm$ 0.013	0.753 $\pm$ 0.053
50	0.961 $\pm$ 0.010	0.806 $\pm$ 0.095	0.973 $\pm$ 0.013	0.755 $\pm$ 0.050
200	0.960 $\pm$ 0.011	0.807 $\pm$ 0.088	0.972 $\pm$ 0.014	0.753 $\pm$ 0.049
1000	0.961 $\pm$ 0.011	0.810 $\pm$ 0.084	0.972 $\pm$ 0.013	0.755 $\pm$ 0.049

Table 6: Effect of accelerated sampling steps on vessel segmentation performance over the external XCA dataset.

Method	Params (M)	Steps	VRAM (GB)	Latency (s/frame)	DSC
TDBM	62.6	1000	0.77	20.09	0.755 $\pm$ 0.049
TDBM	62.6	50	0.77	1.00	0.755 $\pm$ 0.050
TDBM	62.6	10	0.77	0.20	0.753 $\pm$ 0.053
pix2pix	54.4	-	0.45	0.04	0.670 $\pm$ 0.061
nnU-Net	20.6	-	0.21	0.04	0.620 $\pm$ 0.201

Table 7: Computational efficiency and segmentation performance on the external XCA dataset.

Dataset	Method	Accuracy	Sensitivity	DSC
DRIVE	nnUNet †	0.953 $\pm$ 0.003	0.739 $\pm$ 0.044	0.731 $\pm$ 0.013
	MedSegX	0.918 $\pm$ 0.005	0.581 $\pm$ 0.043	0.551 $\pm$ 0.030
	Ours	0.955 $\pm$ 0.006	0.789 $\pm$ 0.086	0.739 $\pm$ 0.044
VessMAP	nnUNet †	0.960 $\pm$ 0.020	0.942 $\pm$ 0.040	0.919 $\pm$ 0.037
	MedSegX	0.644 $\pm$ 0.272	0.782 $\pm$ 0.130	0.591 $\pm$ 0.184
	Ours	0.867 $\pm$ 0.076	0.763 $\pm$ 0.133	0.773 $\pm$ 0.075

Table 8: Zero-shot Segmentation Performance Comparison on DRIVE and VessMAP datasets.

\dagger

indicates baselines trained in a supervised manner using labeled data.

Fig. 6: Visualizations of zero-shot vessel segmentation performance on DRIVE and VessMAP datasets.

Explore similar work

VesselSim: learning 3D blood vessel segmentation without expert annotations
May 25, 2026Erin Rainville, Melissa Ananian, Tristan Mirolla +2Synthetic-to-Real Domain Adaptation Synthetic Data Augmentation
Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada · Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada
MoE-based Feature Adapter for Prompt-free Binary Coronary Artery Segmentation in X-ray Angiography
Aug 25, 2026Lin Xi, Yingliang MaCoronary Artery Segmentation Coronary Angiography
University College London, United Kingdom · University of East Anglia, United Kingdom · King’s College London, United Kingdom
VeCAS: Vessel-Focused Contrast-Free Angiogram Synthesis for Vascular Interventions
Aug 24, 2026De-Xing Huang, Chen-Yu Wang, Hao Liang +10Medical Image Synthesis
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, Beijing 101408, China +1

May 25, 2026J/K move · Enter open · S save

VesselSim: learning 3D blood vessel segmentation without expert annotations

Erin Rainville, Melissa Ananian, Tristan Mirolla, Hassan Rivaz+1

Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada · Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada

Synthetic-to-Real Domain Adaptation Synthetic Data Augmentation Image Segmentation Medical Image Segmentation 3D Medical Image Segmentation

Open paper AskPDF arXiv

Blood vessel segmentation is a core task in medical image analysis for the care of vascular diseases and surgical planning, yet the challenges of providing expert vascular annotations pose a major obstacle for the progress of related deep learning techniques. To address this, we propose VesselSim, a two-stage framework for universal 3D blood vessel segmentation that eliminates the need for real annotated data during training. First, we introduce a stochastic, geometry-driven vascular simulation framework that models recursive branching, curvature-controlled growth, and collision-aware topology, followed by domain-randomized intensity synthesis to generate 16,500 anatomically plausible 3D angiographic volumes. Second, a 3D U-Net is trained solely on this synthetic data. To bridge the domain gap from synthetic to real images at inference time, we introduce a test-time adaptation strategy via a self-supervised mask reconstruction decoder, enabling adaptation to unseen clinical scans without prior domain knowledge. We evaluate VesselSim in a zero-shot setting on multiple real-world datasets spanning MR and CT across several anatomical regions, including the brain and kidneys. Despite being trained exclusively on synthetic data, VesselSim achieves performance competitive with state-of-the-art vascular segmentation foundation models. These findings suggest that learning vessel geometry from synthetic tubular structures is effective for robust cross-domain generalization, substantially reducing the reliance on acquired medical imaging data and more importantly, expert annotations.

Abstract

Explore similar work

VesselSim: learning 3D blood vessel segmentation without expert annotations

MoE-based Feature Adapter for Prompt-free Binary Coronary Artery Segmentation in X-ray Angiography

VeCAS: Vessel-Focused Contrast-Free Angiogram Synthesis for Vascular Interventions