Annotation-Free Cardiac Vessel Segmentation via Knowledge Transfer from Retinal Images

Bin Dong; Fan Yang; Fei Yu; Jie Zhao; Li Zhang; Quanzheng Li; Yanjun Gong; Yuxi Li; Zhi Wang

arxiv: 1907.11483 · v1 · pith:PZADNYYNnew · submitted 2019-07-26 · 📡 eess.IV · cs.CV

Annotation-Free Cardiac Vessel Segmentation via Knowledge Transfer from Retinal Images

Fei Yu , Jie Zhao , Yanjun Gong , Zhi Wang , Yuxi Li , Fan Yang , Bin Dong , Quanzheng Li

show 1 more author

Li Zhang

This is my paper

Pith reviewed 2026-05-24 15:19 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords annotation-free segmentationcoronary artery segmentationknowledge transfergenerative adversarial networksretinal fundus imagesdigital subtraction angiographyshape consistency

0 comments

The pith

A GAN transfers retinal vessel segmentation knowledge to coronary arteries without any annotations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces an annotation-free method for segmenting coronary arteries using a shape-consistent generative adversarial network. The SC-GAN learns from publicly available annotated retinal images to generate synthetic coronary angiography images that keep realistic backgrounds while preserving vessel structures. It then segments these images in an end-to-end process. Tested on 1092 digital subtraction angiography images, the approach claims superior accuracy over existing methods. The key idea is that knowledge from one vascular domain can apply to another without new labels.

Core claim

The authors claim that their shape-consistent generative adversarial network enables annotation-free segmentation of coronary arteries by transferring knowledge from annotated fundus images, through end-to-end training that generates synthetic images preserving retinal vascular structures on coronary backgrounds and segments them, with demonstrated high accuracy on 1092 DSA images.

What carries the argument

Shape-consistent generative adversarial network (SC-GAN) for knowledge transfer between retinal and cardiac vessel segmentation.

If this is right

The method eliminates the need for manual annotation of coronary images.
Synthetic images maintain both background realism and vessel shape consistency.
Segmentation performance exceeds that of classic unsupervised and supervised methods on the tested dataset.
End-to-end training allows the generator and segmenter to improve together.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This suggests similar knowledge transfer could work for other medical imaging domains with scarce labels.
The approach might reduce reliance on expert annotations in clinical settings.
Testing on more diverse patient populations could reveal if the transfer holds across variations in imaging equipment.

Load-bearing premise

Vascular structures from retinal images can be accurately transferred to synthetic coronary images without losing the features needed for correct segmentation.

What would settle it

A drop in segmentation accuracy when applying the model to a new set of real coronary angiography images not used in training would falsify the claim of effective knowledge transfer.

Figures

Figures reproduced from arXiv: 1907.11483 by Bin Dong, Fan Yang, Fei Yu, Jie Zhao, Li Zhang, Quanzheng Li, Yanjun Gong, Yuxi Li, Zhi Wang.

**Figure 1.** Figure 1: The illustration of the proposed SC-GAN, an end-to-end approach for coronary artery segmentation requiring no new manual annotations [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: The illustration of Add U-Net, where the input is an average of fundus photography and DSA image. 3 Experiments In this section, to evaluate the effectiveness of our proposed SC-GAN, we compare the segmentation of four methods: 1) Frangi Algorithm Multi-scale Frangi vessel analysis is used to segment coronary arteries. 2) Classic U-Net. A U-Net model is trained using Frangi Algorithm results as the learn… view at source ↗

**Figure 3.** Figure 3: Comparison of SC-GAN and Cycle-GAN.(a) Fundus patches, (b) DSA patches, (c) synthetic images, (d) synthetic labels. Images segmentation: We annotate 30% of the DSA dataset (328 out of 1092 images) and evaluate our proposed model on it [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Examples of different vessel segmentation methods. (a) Original images, (b) Frangi Algorithm, (c) Classic U-Net, (d) Add U-Net, (e) Proposed SC-GAN, (f) Ground truth [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

Segmenting coronary arteries is challenging, as classic unsupervised methods fail to produce satisfactory results and modern supervised learning (deep learning) requires manual annotation which is often time-consuming and can some time be infeasible. To solve this problem, we propose a knowledge transfer based shape-consistent generative adversarial network (SC-GAN), which is an annotation-free approach that uses the knowledge from publicly available annotated fundus dataset to segment coronary arteries. The proposed network is trained in an end-to-end fashion, generating and segmenting synthetic images that maintain the background of coronary angiography and preserve the vascular structures of retinal vessels and coronary arteries. We train and evaluate the proposed model on a dataset of 1092 digital subtraction angiography images, and experiments demonstrate the supreme accuracy of the proposed method on coronary arteries segmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper proposes a SC-GAN to transfer retinal vessel labels to coronary segmentation without annotations, but the abstract gives no metrics or baselines so the claim cannot be checked.

read the letter

The headline here is that the authors try to sidestep annotation costs for coronary artery segmentation by training on synthetic images that borrow vessel shapes from public retinal data while keeping real coronary backgrounds. They call the model SC-GAN and say it reaches supreme accuracy on 1092 DSA images, but the description contains zero numbers, no Dice scores, no unsupervised baselines, and no ablations, so the result stays unevaluable from the text alone. The approach itself is straightforward: generate synthetics with retinal vessels overlaid on coronary backgrounds, train the segmenter end-to-end on those labeled synthetics, then apply it to real images. That setup directly uses an existing annotated dataset to avoid labeling the target domain, which is a practical move for this subfield. The shape-consistency constraint is a sensible engineering step to make the generated vessels look plausible. The main soft spot is the untested transfer assumption. Retinal and coronary vessels differ in branching angles, diameter distributions, curvature, and imaging physics, so the network could easily latch onto source-specific patterns rather than learning domain-invariant vessel features. The stress-test note correctly flags that nothing isolates whether performance would drop if a different source domain replaced the retinal one. Without experiments that measure this invariance or report quantitative results against standard methods, the central claim rests on an unverified step. This work is aimed at researchers already working on domain adaptation or unsupervised vessel segmentation in medical imaging. A reader looking for concrete numbers or reproducible code would get little value until the results are added. I would not bring it to a reading group as written. It does not yet deserve peer review because the evidence gap is too large; the idea could be worth pursuing once the quantitative controls are in place.

Referee Report

3 major / 1 minor

Summary. The paper proposes an annotation-free method called SC-GAN (shape-consistent generative adversarial network) for segmenting coronary arteries in digital subtraction angiography (DSA) images. It transfers segmentation knowledge from publicly available annotated retinal fundus images by generating synthetic images that combine retinal vascular structures with coronary angiography backgrounds, then trains a segmentation model end-to-end on these synthetics before applying it to 1092 real DSA images, claiming supreme accuracy.

Significance. If the cross-domain transfer of vessel segmentation knowledge holds despite differences in branching patterns, scale, curvature, and imaging physics between retinal and coronary domains, the approach could substantially reduce the need for manual annotations in cardiac vessel segmentation tasks where labeled data is scarce. The use of an existing public retinal dataset as the source of labels is a practical strength, but the absence of any reported metrics, baselines, or ablation studies in the provided text prevents assessment of whether the invariance assumption is realized.

major comments (3)

[Abstract] Abstract: The central claim of 'supreme accuracy' on a dataset of 1092 DSA images is asserted without any accompanying metrics, baselines, error bars, ablation results, or quantitative comparisons, rendering the primary experimental claim unevaluable from the manuscript text.
[Abstract] Abstract (method description): The generative process is described as preserving retinal vascular structures while using coronary backgrounds, but no section isolates or tests whether the segmentation head learns morphology-invariant features rather than retinal-specific patterns; domain differences in bifurcation angles, diameter distributions, and X-ray vs. fundus contrast are not addressed with targeted experiments or failure-case analysis.
[Abstract] Abstract: The annotation-free claim rests on successful knowledge transfer from the retinal source domain, yet the text supplies no evidence (e.g., cross-domain validation or replacement of retinal vessels with an alternative source) that performance on real coronary data would not degrade under changes to the source vessel statistics.

minor comments (1)

[Abstract] Abstract contains a minor typo: 'can some time be infeasible' should read 'can sometimes be infeasible'.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and the evaluation of the knowledge transfer approach. We agree that the abstract requires quantitative support and will revise it accordingly. We address each major comment below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim of 'supreme accuracy' on a dataset of 1092 DSA images is asserted without any accompanying metrics, baselines, error bars, ablation results, or quantitative comparisons, rendering the primary experimental claim unevaluable from the manuscript text.

Authors: We agree that the abstract should be self-contained and include key quantitative results. The full manuscript contains detailed experiments with Dice scores, sensitivity, and comparisons to baselines on the 1092 DSA images. We will revise the abstract to report these metrics explicitly. revision: yes
Referee: [Abstract] Abstract (method description): The generative process is described as preserving retinal vascular structures while using coronary backgrounds, but no section isolates or tests whether the segmentation head learns morphology-invariant features rather than retinal-specific patterns; domain differences in bifurcation angles, diameter distributions, and X-ray vs. fundus contrast are not addressed with targeted experiments or failure-case analysis.

Authors: The shape-consistency constraint in SC-GAN is designed to promote learning of morphology-invariant vessel features. While the original manuscript does not contain a dedicated ablation isolating invariant vs. domain-specific features or explicit failure-case analysis on bifurcation angles and contrast differences, the end-to-end results on real DSA images indicate effective transfer. We will add a discussion of these domain differences and limitations in the revised version. revision: partial
Referee: [Abstract] Abstract: The annotation-free claim rests on successful knowledge transfer from the retinal source domain, yet the text supplies no evidence (e.g., cross-domain validation or replacement of retinal vessels with an alternative source) that performance on real coronary data would not degrade under changes to the source vessel statistics.

Authors: Retinal images were selected as the source due to public availability and shared tubular branching structure with coronary vessels. Performance on the target DSA domain provides supporting evidence for the transfer. We acknowledge the absence of explicit replacement experiments with alternative sources and will expand the discussion to address potential sensitivity to source vessel statistics. revision: partial

Circularity Check

0 steps flagged

No circularity; method uses external public datasets without internal reduction

full rationale

The paper describes an end-to-end SC-GAN trained on synthetic images that combine retinal vessel structures (with labels from a public fundus dataset) and coronary backgrounds from 1092 DSA images. No equations, parameter fits, or derivations are shown that would make any prediction equivalent to its inputs by construction. The annotation-free claim rests on the external retinal dataset and standard GAN training rather than any self-citation chain or self-definitional step. This is a self-contained empirical method description with no load-bearing internal circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no equations or implementation details, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5677 in / 1098 out tokens · 18255 ms · 2026-05-24T15:19:25.551495+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 3 internal anchors

[1]

IEEE Transactions on pattern analysis and machine intelligence 16(6), 641–647 (1994)

Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641–647 (1994)

work page 1994
[2]

In: International conference on medical image computing and computer-assisted intervention

Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A.: Multiscale vessel en- hancement ﬁltering. In: International conference on medical image computing and computer-assisted intervention. pp. 130–137. Springer (1998)

work page 1998
[3]

The Journal of Machine Learning Research 17(1), 2096–2030 (2016)

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The Journal of Machine Learning Research 17(1), 2096–2030 (2016)

work page 2096
[4]

IEEE transactions on medical imaging 22(8), 951–958 (2003)

Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE transactions on medical imaging 22(8), 951–958 (2003)

work page 2003
[5]

In: International conference on med- ical image computing and computer-assisted intervention

Hu, Q., Abr` amoﬀ, M.D., Garvin, M.K.: Automated separation of binary overlap- ping trees in low-contrast color retinal images. In: International conference on med- ical image computing and computer-assisted intervention. pp. 436–443. Springer (2013)

work page 2013
[6]

In: International Conference on Machine Learning

Ioﬀe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning. pp. 448–456 (2015)

work page 2015
[7]

In: International conference on information processing in medical imaging

Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V., Simpson, J., Kane, A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., et al.: Unsupervised domain adap- tation in brain lesion segmentation with adversarial networks. In: International conference on information processing in medical imaging. pp. 597–609. Springer (2017)

work page 2017
[8]

Adam: A Method for Stochastic Optimization

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[9]

IEEE transactions on pattern anal- ysis and machine intelligence 40(7), 1533–1554 (2018)

Lapin, M., Hein, M., Schiele, B.: Analysis and optimization of loss functions for multiclass, top-k, and multilabel classiﬁcation. IEEE transactions on pattern anal- ysis and machine intelligence 40(7), 1533–1554 (2018)

work page 2018
[10]

In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

Nie, D., Trullo, R., Lian, J., Petitjean, C., Ruan, S., Wang, Q., Shen, D.: Medical image synthesis with context-aware generative adversarial networks. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 417–425. Springer (2017) Annotation-Free Cardiac Vessel Segmentation 9

work page 2017
[11]

Journal of computational physics 79(1), 12–49 (1988)

Osher, S., Sethian, J.A.: Fronts propagating with curvature-dependent speed: al- gorithms based on hamilton-jacobi formulations. Journal of computational physics 79(1), 12–49 (1988)

work page 1988
[12]

IEEE transactions on medical imaging 23(4), 501–509 (2004)

Staal, J., Abr` amoﬀ, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501–509 (2004)

work page 2004
[13]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7472–7481 (2018)

work page 2018
[14]

Instance Normalization: The Missing Ingredient for Fast Stylization

Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing in- gredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[15]

mixup: Beyond Empirical Risk Minimization

Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[16]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Zhang, Z., Yang, L., Zheng, Y.: Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9242–9251 (2018)

work page 2018
[17]

In: Proceedings of the IEEE Interna- tional Conference on Computer Vision

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE Interna- tional Conference on Computer Vision. pp. 2223–2232 (2017)

work page 2017
[18]

In: Graphics gems IV

Zuiderveld, K.: Contrast limited adaptive histogram equalization. In: Graphics gems IV. pp. 474–485. Academic Press Professional, Inc. (1994)

work page 1994

[1] [1]

IEEE Transactions on pattern analysis and machine intelligence 16(6), 641–647 (1994)

Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641–647 (1994)

work page 1994

[2] [2]

In: International conference on medical image computing and computer-assisted intervention

Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A.: Multiscale vessel en- hancement ﬁltering. In: International conference on medical image computing and computer-assisted intervention. pp. 130–137. Springer (1998)

work page 1998

[3] [3]

The Journal of Machine Learning Research 17(1), 2096–2030 (2016)

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The Journal of Machine Learning Research 17(1), 2096–2030 (2016)

work page 2096

[4] [4]

IEEE transactions on medical imaging 22(8), 951–958 (2003)

Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE transactions on medical imaging 22(8), 951–958 (2003)

work page 2003

[5] [5]

In: International conference on med- ical image computing and computer-assisted intervention

Hu, Q., Abr` amoﬀ, M.D., Garvin, M.K.: Automated separation of binary overlap- ping trees in low-contrast color retinal images. In: International conference on med- ical image computing and computer-assisted intervention. pp. 436–443. Springer (2013)

work page 2013

[6] [6]

In: International Conference on Machine Learning

Ioﬀe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning. pp. 448–456 (2015)

work page 2015

[7] [7]

In: International conference on information processing in medical imaging

Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V., Simpson, J., Kane, A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., et al.: Unsupervised domain adap- tation in brain lesion segmentation with adversarial networks. In: International conference on information processing in medical imaging. pp. 597–609. Springer (2017)

work page 2017

[8] [8]

Adam: A Method for Stochastic Optimization

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014

[9] [9]

IEEE transactions on pattern anal- ysis and machine intelligence 40(7), 1533–1554 (2018)

Lapin, M., Hein, M., Schiele, B.: Analysis and optimization of loss functions for multiclass, top-k, and multilabel classiﬁcation. IEEE transactions on pattern anal- ysis and machine intelligence 40(7), 1533–1554 (2018)

work page 2018

[10] [10]

In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

Nie, D., Trullo, R., Lian, J., Petitjean, C., Ruan, S., Wang, Q., Shen, D.: Medical image synthesis with context-aware generative adversarial networks. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 417–425. Springer (2017) Annotation-Free Cardiac Vessel Segmentation 9

work page 2017

[11] [11]

Journal of computational physics 79(1), 12–49 (1988)

Osher, S., Sethian, J.A.: Fronts propagating with curvature-dependent speed: al- gorithms based on hamilton-jacobi formulations. Journal of computational physics 79(1), 12–49 (1988)

work page 1988

[12] [12]

IEEE transactions on medical imaging 23(4), 501–509 (2004)

Staal, J., Abr` amoﬀ, M.D., Niemeijer, M., Viergever, M.A., Van Ginneken, B.: Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging 23(4), 501–509 (2004)

work page 2004

[13] [13]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7472–7481 (2018)

work page 2018

[14] [14]

Instance Normalization: The Missing Ingredient for Fast Stylization

Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing in- gredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[15] [15]

mixup: Beyond Empirical Risk Minimization

Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017

[16] [16]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Zhang, Z., Yang, L., Zheng, Y.: Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9242–9251 (2018)

work page 2018

[17] [17]

In: Proceedings of the IEEE Interna- tional Conference on Computer Vision

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE Interna- tional Conference on Computer Vision. pp. 2223–2232 (2017)

work page 2017

[18] [18]

In: Graphics gems IV

Zuiderveld, K.: Contrast limited adaptive histogram equalization. In: Graphics gems IV. pp. 474–485. Academic Press Professional, Inc. (1994)

work page 1994