Vascular anatomy-aware self-supervised pre-training for X-ray angiogram analysis

De-Xing Huang; Chaohui Yu; Xiao-Hu Zhou; Tian-Yu Xiang; Qin-Yi Zhang; Mei-Jiang Gui; Rui-Ze Ma; Chen-Yu Wang

cs.CVFeb 12, 2026

Vascular anatomy-aware self-supervised pre-training for X-ray angiogram analysis

Authors: De-Xing Huang, Chaohui Yu, Xiao-Hu Zhou, Tian-Yu Xiang, Qin-Yi Zhang, Mei-Jiang Gui, Rui-Ze Ma, Chen-Yu Wang,

Organizations: 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 · DAMO Academy, Alibaba Group, Hangzhou 310023, China · Hupan Lab, Hangzhou 310023, China ·

Abstract

Objective: X-ray angiography is the gold standard imaging modality for cardiovascular diseases. However, current deep learning approaches for X-ray angiogram analysis are severely constrained by the scarcity of annotated data. While large-scale self-supervised learning (SSL) has emerged as a promising solution, its potential in this domain remains largely unexplored, primarily due to the lack of effective SSL frameworks and large-scale datasets. Methods: To bridge this gap, we introduce a vascular anatomy-aware masked image modeling (VasoMIM) framework that explicitly integrates domain-specific anatomical knowledge. Specifically, VasoMIM comprises two key designs: an anatomy-guided masking strategy and an anatomical consistency loss. The former strategically masks vessel-containing patches to compel the model to learn robust vascular semantics, while the latter preserves structural consistency of vessels between the original and reconstructed images, enhancing the discriminability of the learned representations. In conjunction with VasoMIM, we curate XA-170K, the largest X-ray angiogram pre-training dataset to date. Results: We validate VasoMIM on four downstream tasks across six datasets, where it demonstrates superior transferability and achieves state-of-the-art performance compared to existing methods. Conclusion: Incorporating vascular anatomical knowledge into self-supervised pre-training enables more discriminative and transferable representations for X-ray angiogram analysis. Significance: These findings highlight the significant potential of VasoMIM as a foundation model for advancing a wide range of X-ray angiogram analysis tasks. VasoMIM and XA-170K are available on GitHub and Hugging Face, respectively.

Figures & tables

Fig. 1 : Conceptual comparison. (a) Conventional MIM masks patches based on generic rules and utilizes standard supervision to reconstruct the original image. (b) VasoMIM employs anatomical guidance to selectively mask vessel-relevant regions (highlighted in yellow) and enforces structural consistency via anatomical supervision , enabling the model to learn richer vascular representations.

Fig. 2 : Generic vs. anatomy-guided masking strategies. While generic methods like AMT [ 11 ] and HPM [ 12 ] fail to prioritize vascular structures, VasoMIM effectively focuses on vessel-relevant regions. Red denotes a higher masking probability, while blue indicates the opposite.

Aspect	VasoMIM-v1 [ 13 ]	VasoMIM	Main Contribution
Anatomical Guidance	Instance-level guidance from Frangi filter	Incorporating population-level guidance learned from large-scale X-ray angiograms	More reliable anatomical guidance by complementing handcrafted vesselness cues with learned anatomical knowledge
Pre-training Dataset	$20$ K X-ray angiograms	$170$ K X-ray angiograms	The largest publicly available dataset for X-ray angiogram pre-training
Downstream Validation	$1$ task (vessel segmentation) on $3$ datasets	$4$ tasks (segmentation and detection) on $6$ datasets	More comprehensive evaluation across diverse X-ray angiogram analysis tasks
Experimental Analysis	Basic ablation studies	Extended ablations and analyses of scalability, label efficiency, pre-training datasets, and robustness	More in-depth understanding of the effectiveness and generalizability of VasoMIM

TABLE I : Comparison between the preliminary conference version VasoMIM-v1 [ 13 ] and this extended version.

Fig. 3 : Overview of VasoMIM. Given an input X-ray angiogram, VasoMIM constructs anatomical guidance from two complementary sources: instance-level vesselness cues extracted by Frangi filter and population-level anatomical knowledge learned by a lightweight segmentor from the large-scale XA- $170$ K. Based on the co-guidance from these two sources, a patch-wise vascular anatomical distribution $f(g_{i})$ is computed to guide the masking strategy, thereby prioritizing vessel-relevant regions during pre-training. The masked image is then reconstructed by an encoder-decoder model. Finally, the model is optimized by minimizing the total objective $\mathcal{L}_{\rm MIM}$ , which combines the standard pixel-level reconstruction loss $\mathcal{L}_{\rm rec.}$ and the anatomical consistency loss $\mathcal{L}_{\rm cons.}$ to learn discriminative vascular representations.

Usage	Dataset	#train	#test
Pre-training	CADICA [ 47 ]	$6,594$	$-$
	SYNTAX [ 48 ]	$2,943$	$-$
	XCAD [ 49 ]	$1,621$	$-$
	CoronaryDominance [ 50 ]	$160,320$	$-$
	Total	$171,478$	$-$
Segmentation	ARCADE-V [ 51 ]	$1,000$	$300$

TABLE II : Statistics of XA-

170

K and downstream datasets.

Fig. 4 : Overview of downstream tasks. We evaluate VasoMIM on four distinct tasks: vessel segmentation, vessel segment segmentation, stenosis segmentation, and stenosis detection.

Method	Pre-training Data	ARCADE-V		CAXF		XACV		ARCADE-S	ARCADE-VS	Avg. Rank
Method	Pre-training Data	DSC (%)	clDice (%)	DSC (%)	clDice (%)	DSC (%)	clDice (%)	DSC (%)	DSC (%)	Avg. Rank
Traditional
Frangi Filter [ 19 ] [ MICCAI ’98]	$-$	$41.30$	$40.91$	$64.01$	$65.73$	$58.46$	$57.15$	$-$	$-$	$-$
From Scratch
UNet [ 37 ] [ MICCAI ’15]	$-$	$71.44$ $\pm 0.32$	$70.67$ $\pm 0.40$	$82.76$ $\pm 0.75$	$84.95$ $\pm 0.56$	$78.18$ $\pm 0.34$	$74.71$ $\pm 0.44$	$27.04$ $\pm 3.65$	$38.77$ $\pm 0.82$	$23.00$ $\pm 0.93$
nnUNet [ 53 ] [ Nat. Methods ’21]	$-$	$75.85$ $\pm 0.40$	$77.82$ $\pm 0.16$	$87.26$ $\pm 0.18$	$90.31$ $\pm 0.26$	$84.05$ $\pm 0.17$	$82.63$ $\pm 0.15$	$43.14$ $\pm 1.31$	$47.86$ $\pm 2.37$	$13.88$ $\pm 4.55$

TABLE III : Quantitative comparison on segmentation tasks. We report “

{\rm mean}\pm{\rm std}

” over five random seeds for all learning-based methods. Since Frangi filter is deterministic, we report only “

{\rm mean}

”. The best results are highlighted in bold , and the second-best are underlined . Paired

t

-tests are conducted between VasoMIM-v1 and VasoMIM based on five seed-level results.

\ast

p<0.05

\ast\ast

p<0.01

\ast\ast\ast

p<0.001

Panel of Figure 8 — Fig. 5 : Qualitative analysis of the anatomy-guided masking strategy. Left: The proportion of vessel-containing patches among all masked patches across training epochs. Right: Visual comparison. We highlight failure cases of Frangi filter where UNeXt-S [ 20 ] , leveraging a strong inductive bias, extracts vascular anatomy more accurately to provide precise guidance. The heatmaps display the patch-wise masking ratio, i.e. , $\frac{1}{E}\sum_{j=1}^{E}\mathbb{I}\left(\text{Patch$ x_{i} $is masked in epoch$ j $}\right)$ .

Method	Pre-training Data	mAP $50$ (%)	mAP $75$ (%)	mAP (%)	Avg. Rank
From Scratch
Faster R-CNN [ 38 ] [ TPAMI ’16]	$-$	$88.37$ $\pm 0.90$	$19.01$ $\pm 1.77$	$36.63$ $\pm 0.63$	$20.33$ $\pm 0.58$
DETR [ 67 ] [ ECCV ’20]	$-$	$89.30$ $\pm 0.79$	$19.35$ $\pm 2.00$	$36.51$ $\pm 0.57$	$19.00$ $\pm 2.65$
General SSL
MoCo v3 [ 24 ] [ ICCV ’21]	XA- $170$ K	$91.54$ $\pm 1.04$	$21.74$ $\pm 0.67$	$38.74$ $\pm 0.32$	$14.00$ $\pm 1.00$
DINO [ 25 ] [ ICCV ’21]	XA- $170$ K	$88.83$ $\pm 0.46$	$21.12$ $\pm 0.78$	$37.65$ $\pm 0.41$	$17.00$ $\pm 1.00$

TABLE IV : Quantitative results for the stenosis detection task. We report “

{\rm mean}\pm{\rm std}

” over five random seeds. The best and second-best results are highlighted in bold and underlined , respectively. Paired

t

-tests are conducted between VasoMIM-v1 and VasoMIM based on five seed-level results.

\ast

p<0.05

\ast\ast

p<0.01

{\rm n.s.}

: not significant.

Guidance	$\mathcal{L}_{\rm cons.}$	ARCADE-V		XACV
Guidance	$\mathcal{L}_{\rm cons.}$	DSC (%)	clDice (%)	DSC (%)	clDice (%)
$-$	$-$	$79.31$ $\pm 0.16$ ∗∗∗	$81.05$ $\pm 0.25$ ∗∗∗	$84.52$ $\pm 2.12$ n.s.	$82.25$ $\pm 2.32$ n.s.
$-$	✓	$79.85$ $\pm 0.14$ ∗	$\underline{81.59}$ $\pm 0.16$ ∗	$85.79$ $\pm 0.52$ n.s.	$83.66$ $\pm 0.77$ n.s.
$-$	✓	$\uparrow 0.54$	$\uparrow 0.54$	$\uparrow 1.27$	$\uparrow 1.41$
✓	$-$	$\underline{79.87}$ $\pm 0.09$ ∗∗	$81.41$ $\pm 0.11$ ∗∗	$\underline{85.92}$ $\pm 0.04$ ∗	$\underline{83.78}$ $\pm 0.15$ ∗
✓	$-$	$\uparrow 0.56$	$\uparrow 0.36$	$\uparrow 1.40$	$\uparrow 1.53$

TABLE V : Ablation study of the anatomy-guided masking strategy and the anatomical consistency loss. Paired

t

-tests are conducted between each variant and VasoMIM using five seed-level results.

\ast

p<0.05

\ast\ast

p<0.01

\ast\ast\ast

p<0.001

{\rm n.s.}

: not significant.

Baseline
Loss	Params. (M)	FLOPs (G)	ARCADE-V	XACV
$-$	$111.91$	$19.01$	$79.87$ $\pm 0.09$	$85.92$ $\pm 0.04$
Feature Distillation
DINO [ 25 ] [ ICCV ’21]	$197.95$	$52.90$	$79.81$ $\pm 0.12$	$85.03$ $\pm 2.27$
DINO [ 25 ] [ ICCV ’21]	$+86.04$	$+33.89$	$\downarrow 0.06$	$\downarrow 0.89$
CLIP [ 68 ] [ ICML ’21]	$169.43$	$41.66$	$79.79$ $\pm 0.13$	$85.29$ $\pm 1.47$

TABLE VI : Comparison with alternative reconstruction objectives. GFLOPs are calculated on an NVIDIA H

20

GPU using a single

224\times 224

masked RGB image. DSC is reported in this table.

Baseline
Loss	SS $\uparrow$	CHI $\uparrow$	DBI $\downarrow$
$-$	$0.27$	$1.39$	$1.37$
Feature Distillation
DINO [ 25 ] [ ICCV ’21]	$0.03$	$0.06$	$6.42$
CLIP [ 68 ] [ ICML ’21]	$0.08$	$0.18$	$3.71$
DINOv3 [ 56 ] [ arXiv ’25]	$\underline{0.29}$	$\underline{1.65}$	$\underline{1.24}$

TABLE VII : Comparison of clustering metrics on XACV training set .

Guidance	ARCADE-V		XACV
Guidance	DSC (%)	clDice (%)	DSC (%)	clDice (%)
$-$	$79.85$ $\pm 0.14$ ∗	$81.59$ $\pm 0.16$ ∗	$85.79$ $\pm 0.52$ n.s.	$83.66$ $\pm 0.77$ n.s.
Frangi-only [ 13 ] [ AAAI ’26]	$79.90$ $\pm 0.16$ ∗∗	$81.57$ $\pm 0.28$ ∗∗	$85.80$ $\pm 0.27$ ∗	$83.75$ $\pm 0.37$ ∗
Frangi-only [ 13 ] [ AAAI ’26]	$\uparrow 0.05$	$\downarrow 0.02$	$\uparrow 0.01$	$\uparrow 0.09$
Segmentor-only	$\underline{79.98}$ $\pm 0.13$ ∗∗∗	$\underline{81.77}$ $\pm 0.19$ ∗∗	$\underline{85.98}$ $\pm 0.08$ ∗∗	$\underline{83.91}$ $\pm 0.13$ ∗
Segmentor-only	$\uparrow 0.13$	$\uparrow 0.18$	$\uparrow 0.19$	$\uparrow 0.25$

TABLE VIII : Comparison between instance-level and population-level anatomical guidance. Paired

t

-tests are conducted between each variant and VasoMIM using five seed-level results.

\ast

p<0.05

\ast\ast

p<0.01

\ast\ast\ast

p<0.001

{\rm n.s.}

: not significant.

Fig. 6 : Qualitative comparison between VasoMIM-v1 [ 13 ] and VasoMIM on vessel segmentation (ARCADE-V), stenosis segmentation (ARCADE-S), vessel segment segmentation (ARCADE-VS), and stenosis detection (Stenosis). Yellow boxes highlight improved structural continuity, lesion delineation, or localization.

Fig. 7 : Hyper-parameter sensitivity of the mixing coefficient

\eta

on ARCADE-V and XACV. A larger

\eta

assigns a higher relative weight to Frangi-derived instance-level guidance, with

\eta=0

and

\eta=1

corresponding to segmentor-only and Frangi-only guidance, respectively. All results are averaged over five random seeds.

Guidance	Format	Sampling	ARCADE-V	XACV
$-$	$-$	Random	$79.85$ $\pm 0.14$ ∗	$85.79$ $\pm 0.52$ n.s.
Co-guidance	Soft	Uniform	$79.99$ $\pm 0.14$ ∗	$85.94$ $\pm 0.06$ ∗
	Soft	Uniform	$\uparrow 0.14$	$\uparrow 0.15$
	Hard	$f(g_{i})$	$\underline{80.03}$ $\pm 0.04$ ∗	$\underline{85.98}$ $\pm 0.03$ ∗
	Hard	$f(g_{i})$	$\uparrow 0.18$	$\uparrow 0.19$
			$\bm{80.25}$ $\pm 0.12$	$\bm{86.09}$ $\pm 0.09$

TABLE IX : Ablation of patch sampling strategy and guidance format. DSC is reported in this table. “Format” refers to the segmentor output used in guidance: “Hard” denotes a binarized mask, “Soft” denotes a probability map, and “

-

” indicates that no segmentor output is used. Paired

t

-tests are conducted between each variant and VasoMIM using five seed-level results.

\ast

p<0.05

{\rm n.s.}

: not significant.

Case	Manner	$\beta_{0}$	$\beta_{E}$	ARCADE-V	XACV
Random	$-$	$0$	$0$	$79.85$ $\pm 0.14$	$85.79$ $\pm 0.52$
Guided	Weak-to-Strong			$\bm{80.25}$ $\pm 0.12$	$\bm{86.09}$ $\pm 0.09$
		$0$	$0.5$	$\uparrow\bm{0.40}$	$\uparrow\bm{0.30}$
		$0$	$0.75$	$79.91$ $\pm 0.13$	$\underline{85.99}$ $\pm 0.07$
		$0$	$0.75$	$\uparrow 0.06$	$\uparrow 0.20$
		$0$	$1$	$\underline{79.96}$ $\pm 0.06$	$85.94$ $\pm 0.13$

TABLE X : Evaluation of different masking strategies. DSC is reported in this table.

Fig. 8 : Data scaling analysis. We evaluate the impact of increasing pre-training data size from

10

K to

170

K on downstream performance. Left: ARCADE-V. Right: XACV.

Fig. 9 : Scalability analysis of VasoMIM. We evaluate the impact of scaling the model from ViT-B to ViT-L and ViT-H on downstream performance. Left: ARCADE-V. Right: XACV.

Fig. 10 : Label-efficient fine-tuning on ARCADE-V. VasoMIM consistently outperforms competing SSL alternatives, notably achieving superior performance with only

10\%

of labels compared with nnUNet [ 53 ] and TransUNet [ 41 ] trained from scratch using

100\%

of labels.

\ast

p<0.05

\ast\ast

p<0.01

\ast\ast\ast

p<0.001

(paired

t

-test).

Method	Pre-training Data	ARCADE-V	XACV
MAE [ 14 ] [ CVPR ’22]	ImageNet- $1$ K	$78.10$ $\pm 0.21$	$83.69$ $\pm 1.48$
	XA- $170$ K	$79.39$ $\pm 0.15$	$84.84$ $\pm 1.79$
	$\Delta$	$\uparrow 1.29$	$\uparrow 1.15$
LocalMIM [ 54 ] [ CVPR ’23]	ImageNet- $1$ K	$77.55$ $\pm 0.66$	$83.00$ $\pm 0.61$
	XA- $170$ K	$78.79$ $\pm 0.20$	$83.18$ $\pm 1.84$
	$\Delta$	$\uparrow 1.24$	$\uparrow 0.18$

TABLE XI : Impact of pre-training dataset. We compare models pre-trained on generic, medical-specific, and our XA-

170

K datasets. DSC is reported in this table.

Fig. 11 : Segmentor-based Pre-training (SPT) vs. MIM. VasoMIM significantly outperforms SPT, achieving DSC improvements of

+3.77\%

on ARCADE-V and

+4.09\%

on XACV.

\ast\ast\ast

p<0.001

(paired

t

-test).

Fig. 12 : Hyper-parameter sensitivity of the masking ratio

\gamma

on ARCADE-V. All results are averaged over five random seeds.

$\alpha$	Method	ARCADE-V		XACV
$\alpha$	Method	DSC (%)	clDice (%)	DSC (%)	clDice (%)
$50$	Frangi	$12.06$	$8.20$	$32.16$	$22.58$
$50$	VasoMIM	$79.71$ $\pm 0.12$	$81.26$ $\pm 0.21$	$86.02$ $\pm 0.08$	$84.05$ $\pm 0.12$
$80$	Frangi	$27.06$	$22.34$	$66.70$	$62.51$
$80$	VasoMIM	$79.67$ $\pm 0.18$	$81.31$ $\pm 0.22$	$86.03$ $\pm 0.06$	$84.03$ $\pm 0.18$
$92$	Frangi	$41.30$	$40.91$	$58.46$	$57.15$

TABLE XII : Sensitivity analysis of VasoMIM to the percentile threshold

\alpha

of Frangi filter.

Filter	ARCADE-V		XACV
Filter	DSC (%)	clDice (%)	DSC (%)	clDice (%)
$-$	$79.31$ $\pm 0.16$	$81.05$ $\pm 0.25$	$84.52$ $\pm 2.12$	$82.25$ $\pm 2.32$
Canny † [ 73 ] [ TPAMI ’86]	$79.32$ $\pm 0.12$	$80.86$ $\pm 0.24$	$85.73$ $\pm 0.06$	$83.60$ $\pm 0.09$
Canny † [ 73 ] [ TPAMI ’86]	$\uparrow 0.01$	$\downarrow 0.19$	$\uparrow 1.21$	$\uparrow 1.35$
Sato ‡ [ 74 ] [ MedIA ’98]	$\underline{79.99}$ $\pm 0.14$	$81.53$ $\pm 0.31$	$85.84$ $\pm 0.09$	$83.75$ $\pm 0.09$
Sato ‡ [ 74 ] [ MedIA ’98]	${\color[rgb]{0,0.707,0}\uparrow 0.68}$	$\uparrow 0.48$	$\uparrow 1.32$	$\uparrow 1.50$

TABLE XIII : Evaluation of different filters.

\dagger

: General edge detector.

\ddagger

: Vesselness filter.

-

denotes the vanilla baseline, i.e., MAE (

\gamma=0.5

), without anatomy-guided masking or anatomical consistency loss.

$\mathcal{L}$	ARCADE-V		XACV
$\mathcal{L}$	DSC (%)	clDice (%)	DSC (%)	clDice (%)
$-$	$79.87$ $\pm 0.09$	$81.41$ $\pm 0.11$	$85.92$ $\pm 0.04$	$83.78$ $\pm 0.15$
L 1	$79.64$ $\pm 0.17$	$81.21$ $\pm 0.24$	$85.87$ $\pm 0.08$	$83.85$ $\pm 0.07$
L 1	$\downarrow 0.23$	$\downarrow 0.20$	$\downarrow 0.05$	$\uparrow 0.07$
Dice	$79.94$ $\pm 0.11$	$81.64$ $\pm 0.19$	$85.93$ $\pm 0.08$	$84.04$ $\pm 0.14$
Dice	$\uparrow 0.07$	$\uparrow 0.23$	$\uparrow 0.01$	$\uparrow 0.26$

TABLE XIV : Evaluation of different metric functions.

Medical Image Representation Learning

Medical image representation learning builds reusable visual features for diagnosis, segmentation, retrieval, and other clinical imaging tasks.

See all 260 papers

Mar 1, 2026

Uncertainty-Aware Concept and Motion Segmentation for Semi-Supervised Angiography Videos

Segmentation of the main coronary artery from X-ray coronary angiography (XCA) sequences is crucial for the diagnosis of coronary artery diseases. However, this task is challenging due to issues such as blurred boundaries, inconsistent radiation contrast, complex motion patterns, and a lack of annotated images for training. Although Semi-Supervised Learning (SSL) can alleviate the annotation burden, conventional methods struggle with complicated temporal dynamics and unreliable uncertainty quantification. To address these challenges, we propose SAM3-based Teacher-student framework with Motion-Aware consistency and Progressive Confidence Regularization (SMART), a semi-supervised vessel segmentation approach for X-ray angiography videos. First, our method utilizes SAM3's unique promptable concept segmentation design and innovates a SAM3-based teacher-student framework to maximize the performance potential of both the teacher and the student. Second, we enhance segmentation by integrating the vessel mask warping technique and motion consistency loss to model complex vessel dynamics. To address the issue of unreliable teacher predictions caused by blurred boundaries and minimal contrast, we further propose a progressive confidence-aware consistency regularization to mitigate the risk of unreliable outputs. Extensive experiments on three datasets of XCA sequences from different institutions demonstrate that SMART achieves state-of-the-art performance while requiring significantly fewer annotations, making it particularly valuable for real-world clinical applications where labeled data is scarce. Our code is available at: https://github.com/qimingfan10/SMART.

Yu Luo, Guangyu Wei, Yangfan Li +2

School of Mathematical Sciences, Ocean University of China, China · School of Haide, Ocean University of China, China · School of Intelligent Science and Technology, Nanjing University, China +2

Semi-Supervised Medical Image Segmentation Coronary Artery Segmentation Coronary Angiography Teacher-Student Learning Segmentation Uncertainty Estimation

Ask PDF

Apr 8, 2026

VAMAE: Vessel-Aware Masked Autoencoders for OCT Angiography

Optical coherence tomography angiography (OCTA) provides non-invasive visualization of retinal microvasculature, but learning robust representations remains challenging due to sparse vessel structures and strong topological constraints. Many existing self-supervised learning approaches, including masked autoencoders, are primarily designed for dense natural images and rely on uniform masking and pixel-level reconstruction, which may inadequately capture vascular geometry. We propose VAMAE, a vessel-aware masked autoencoding framework for self-supervised pretraining on OCTA images. The approach incorporates anatomically informed masking that emphasizes vessel-rich regions using vesselness and skeleton-based cues, encouraging the model to focus on vascular connectivity and branching patterns. In addition, the pretraining objective includes reconstructing multiple complementary targets, enabling the model to capture appearance, structural, and topological information. We evaluate the proposed pretraining strategy on the OCTA-500 benchmark for several vessel segmentation tasks under varying levels of supervision. The results indicate that vessel-aware masking and multi-target reconstruction provide consistent improvements over standard masked autoencoding baselines, particularly in limited-label settings, suggesting the potential of geometry-aware self-supervised learning for OCTA analysis.

Ilerioluwakiiye Abolade, Prince Mireku, Kelechi Chibundu +4

Federal University of Agriculture Abeokuta, Nigeria · ML Collective · Ashesi University, Ghana +1

Masked Autoencoders Self-Supervised Pre-Training Medical Imaging Retinal Vessel Segmentation

Ask PDF

May 25, 2026

VesselSim: learning 3D blood vessel segmentation without expert annotations

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.

Erin Rainville, Melissa Ananian, Tristan Mirolla +2

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

Ask PDF

Abstract

Medical Image Representation Learning

Explore similar work

Uncertainty-Aware Concept and Motion Segmentation for Semi-Supervised Angiography Videos

VAMAE: Vessel-Aware Masked Autoencoders for OCT Angiography

VesselSim: learning 3D blood vessel segmentation without expert annotations