List of Papers Browse by Subject Areas Author List
Abstract
Precise analysis of ophthalmic diseases necessitates accurate vessel segmentation, image registration, and bifurcation detection across multi-modal images. However, automating these tasks is hindered by expensive manual annotation, which is further exacerbated by various imaging modalities and diseases. Existing generative methods for reducing labeling burden primarily struggle with two challenges: (1) reliance on target-domain image-label pairs and (2) difficulties in achieving a proper balance between controllability and appearance realism. In addition, there is a lack of effective methods for generating modality-dependent labels that can support multiple tasks. In this paper, we propose a unified framework that first generates modality-specific vessel masks and bifurcation labels, followed by the synthesis of realistic target-modality images that are structurally consistent with the generated labels, for zero-shot multi-modal and multi-task fundus image analysis. Specifically, we first rely on vascular dynamics simulations to design a flexible pipeline to synthesize labels. Then, by disentangling structure and appearance via phase maps during the generation, we are able to generate realistic images without requiring image-label pairs from target modalities. Our framework is evaluated under four paradigms across three tasks, including segmentation, registration, and keypoint detection, where our method achieves strong zero-shot performance. Our code is available at: https://github.com/kaiwenli325/UniVessel.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1262_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/kaiwenli325/UniVessel
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiKai_Crossdomain_MICCAI2026,
author = { Li, Kaiwen AND He, Hangzhou AND Hu, Jiakui AND Yao, Zhengjian AND Zeng, Shuang AND Fu, Ourui AND Zhu, Lei AND Lu, Yanye},
title = { { Cross-domain Controllable Generation Enables Zero-shot Fundus Image Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contributions of this study are as follows:
-
The paper’s primary contribution is the proposal of UniVessel, a unified framework for zero-shot fundus image analysis that generates modality-specific vessel masks and bifurcation labels, and then synthesizes realistic target-modality images that remain structurally consistent with those labels.
-
The paper also introduces a phase-based generation strategy, where local phase maps are used as an intermediate bridge to decouple vessel structure from modality-specific appearance, enabling cross-domain controllable generation without requiring target-domain image-label pairs.
-
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The paper proposes a phase-based generation strategy using local phase maps as a bridge between vessel structure and modality-specific appearance. This approach effectively enables controllable cross-domain generation while reducing interference from modality-specific intensity and contrast variations.
-
The paper proposes a modality-aware label generation pipeline that synthesizes vessel masks and bifurcation labels without directly reusing masks from other datasets.
-
The paper presents a strong empirical evaluation across multiple downstream tasks, including segmentation, registration, and bifurcation detection.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The analysis of the proposed label-generation pipeline is somewhat limited: Although the paper includes ablations on the use of phase maps and the drop mechanism, it does not isolate the contribution of the generated masks themselves—for example, by comparing the proposed mask generator against using training masks under the same phase-based image-generation framework. As a result, it is difficult to determine how much of the final improvement comes from the phase-based image synthesis and how much comes from the proposed mask generation pipeline.
-
The design choices in the label generator are not sufficiently analyzed: In particular, the paper states that the vessel diameter distribution is controlled by the parameters γ, β, and k, but does not provide a sensitivity analysis or justification for these choices. Since these parameters may substantially affect the morphology of the generated vessel masks, a more detailed study would make the proposed framework more convincing.
-
The direct validation of the synthesized vessel structures is limited: Given the zero-shot setting, a pixel-wise comparison against target-domain ground-truth labels may not be feasible. However, since a key claim of the paper is that UniVessel-label generates modality-aware vascular structures, it would strengthen the paper to provide additional evidence that these synthesized structures match the target modality in terms of morphology and topology, for example through statistics of vessel density, tortuosity, branching patterns, diameter distribution, or expert-based qualitative assessment.
-
The empirical comparison could be strengthened with more relevant baselines and clearer experimental settings: While the paper compares against CUT, CycleGAN, DDPM, and Rectified Flow, it does not include enough direct comparisons with more fundus-specific or more controllable generation methods discussed in the related work. In addition, given the large performance gaps reported in the tables, the paper would benefit from clearer descriptions of the training and evaluation settings used for the compared baselines, in order to better assess the fairness of the comparison.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-motivated framework that combines phase-based cross-domain image generation with modality-aware label synthesis. However, much of the evidence relies on downstream performance, while the individual components of the framework are not analyzed in sufficient depth.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The main contribution of this work is a unified zero-shot framework for fundus image analysis, combining modality-aware label synthesis with phase map as intermediate for controllable image generation, and showing that the generated pairs are useful for segmentation, registration, and bifurcation detection
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The major strength of the paper lies in its modality-specific label generation. The modality-specific label generation effectively incorporates medical priors and addresses not only the issue that the modality-dependent nature of vessel masks can amplify discrepancies between generated and real images, but also the potential domain-related issues.
Another strength is the overall completeness and coherence of the framework. The pipeline first performs modality-aware label generation, then phase-mediated image generation, and finally uses the generated data for segmentation, registration, and bifurcation detection. This makes the framework logically well structured and practically meaningful. Its value lies not only in whether the generated images look realistic, but also in whether the synthesized data can genuinely support downstream training. In this sense, although the work does not introduce an entirely new problem, it presents a more complete framework for cross-modal zero-shot fundus analysis by building on existing ideas in vessel synthesis and controllable generation.
These strengths are further supported by the experimental results. The authors evaluate the generated data both qualitatively and quantitatively through downstream performance on vessel segmentation, image registration, and bifurcation detection, with comparisons against several generative baselines, task-specific baselines, and ablation settings. In particular, phase map is not novel itself, but the ablation studies show using phase map is effective. This not only highlights phase map is an effective design choice, but also validates the overall framework design.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main weakness is that the technical novelty is moderate rather than strong. Vessel structure synthesis is not new, mask-guided image generation is not new, and phase maps are not new either. The main novelty lies in how these components are combined, especially through the phase-mediated generation formulation. As a result, the paper is stronger as a system-level and problem-driven contribution than as a highly original algorithmic contribution.
Another weakness is that the claim of “completely zero-shot” is not fully rigorous. The method does not require paired target-domain image-label annotations, but it still uses target-domain images during training. Therefore, the setting is closer to label-free target-domain adaptation or target-domain supervision without labels than to a strictly target-free zero-shot setting, and the wording may overstate the claim.
In addition, UniVessel-label is essentially a rule-based label synthesis pipeline built on hand-crafted priors such as vessel graph simulation, Murray’s law, spline interpolation, and manually controlled geometric parameters. While this makes the process controllable, it also raises questions about how closely the synthesized vessel structures match the true anatomical and statistical distribution of real target-domain data, especially in more complex pathological settings.
Finally, although the paper evaluates multiple downstream tasks, they are all vessel-centric tasks. This means the “multi-task” claim is valid, but still limited in scope. The paper also provides only limited analysis of why the synthesized data are anatomically realistic beyond downstream task performance, and stronger evidence on structural realism and distributional fidelity would make the conclusions more convincing.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, this is a well-motivated and technically coherent MICCAI paper. Its motivation is not simply to generate realistic fundus images, but to address a practical supervision bottleneck in cross-modal fundus analysis, where target modalities often lack fine-grained annotations and existing mask-guided methods suffer from modality-dependent structural mismatch. Its main value lies less in a highly novel individual component than in demonstrating an effective pipeline for cross-modal zero-shot fundus analysis. While the technical novelty is moderate, the problem is meaningful, the framework is complete, and the experimental validation is solid.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes a generative framework for fundus image-mask synthesis. The major contributions are the introduction of phase maps as an intermediate modality and the decoupled generation paradigm. The proposed framework can be trained without paired target images and masks, mitigating the scarcity of medical data. The proposed method is evaluated on three downstream tasks, proving its superior performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) The paper explains the motivation of the proposed framework clearly. A generative method that can be trained without target data pairs is indeed advantageous. The decoupling of structures and styles is not a new notion, but it is reasonable to apply it in this scenario. The introduction of phase maps as an intermediate modality is intuitive, as they naturally filter out many style-related components of images. 2) The experiments are very comprehensive. The selected three downstream tasks are closely related to vascular image analysis. Therefore, it is convincing to evaluate the methods on them. The improvements of the proposed framework are significant compared to competing approaches.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) The paper mentions several times that the proposed framework has the advantage of generating modality-specific masks. On page 5, the generation of modality-specific masks is achieved by tuning the parameters of UniVessel-label. It is a bit unclear how to tune these parameters. Wouldn’t imaging properties, such as FOVs, be different even for the same target modality on different devices? How much expertise does it require to tune these parameters? These factors will affect the clinical usability of the proposed method. 2) One of the key contributions of the paper is the utilization of phase maps. The (a) and (b) parts of UniVessel-img can be merged into one model if not using phase maps. Therefore, it would be crucial to conduct an experiment under this model configuration to verify the significance of these phase maps. However, there is no such experiment in the current ablation study, which weakens the motivation for using the phase maps. Besides, the setting of Canny in the ablation study is not explained. It seems that the complete model does not have such a component. Is this the well-known Canny detector? Such information missing causes confusion and an incomplete presentation. 3) There are no image-level metrics, such as FID, presented to quantitatively evaluate the quality of the synthesized images. 4) The performance of DDPM and Rectified Flow in the experiments is significantly worse. During the training process of the proposed framework, the paired public data and the image-only target data are both visible to the model. Are DDPM and Rectified Flow trained in the same way or additionally with the target masks? If it is the first case, the convincingness of the relevant experiments is decreased. This is because training a conditional DDPM for image-mask generation requires paired images and masks, and deviation from such a training paradigm will no doubt lead to significant performance degradation. Although this incomplete data usage is aligned with the proposed method and therefore seems to be fair, the resulting huge improvement of the proposed approach can only suggest its qualitative improvement over previous methods, which is the achievement of a paired-data-free generation paradigm. The quantitative improvement itself is not that informative, as the strongly degraded performance of DDPM and Rectified Flow is not their upper bound.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed generative framework has a clear architecture design and shows potential for clinical usage. The introduction of phase maps as an intermediate modality is impressive, and the experiments are comprehensive and relatively convincing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their constructive feedback. We address the main concerns below: R2-Q1 (generated vs real masks): We agree that this comparison would help isolate UniVessel-label. However, our focus is the target-label-free setting, where real target masks are unavailable and UniVessel-label provides modality-aware generation. Also, public masks may introduce a structural domain gap. We will consider this comparison in future work. R2-Q2 & R4-Q1 (UniVessel-label parameters): For each target modality, instead of per-image tuning, we determine modality-level parameters by visually comparing generated masks with a few unlabeled target images. Then, we introduce small random perturbations to cover intra-modality variations. Since UniVessel-label is computationally efficient, this process only takes a few minutes for a single modality. We will clarify it and a more systematic sensitivity analysis is left for future work. R2-Q3 & R3-Q3 & R3-Q4 & R4-Q3 (analysis of generated masks/images): We appreciate the reviewers’ suggestion for more direct evidence on the realism of generated masks and images. The current manuscript provides complementary evidence. Fig. 3 qualitatively shows that the generated masks and images follow modality-specific vessel structures and target styles. In addition, Tables 1–3 show strong performance on three vessel-dependent tasks without any real target image-label pair, which provides task-relevant evidence that the generated vessels and images are structurally meaningful and distributionally relevant. Notably, the LSCI, FP, and FIRE datasets contain pathological variations, and the consistent performance suggests practical generalization to clinical data. We agree that downstream performance cannot fully replace direct evaluation, including for data with pathological variations. Following the rebuttal policy, we will consider more systematic structural validation as future work and clarify the vessel-centric scope of the “multi-task” claim in the final version. R2-Q4 & R4-Q4 (experiment comparisons): DDPM and Rectified Flow are trained with the same data and setting as UniVessel: no target masks are used. We acknowledge that their current results are not supervised upper bounds when paired target data are available. The experiments are intended to demonstrate that, under the target-mask-free setting, UniVessel can achieve target appearance realism and controllable vessel structures, thereby reducing reliance on costly annotations. Moreover, some prior works relying on manually annotated target image-label pairs cannot be applied to the label-free scenario. We will clarify the training settings of the baselines. R3-Q1 (novelty): UniVessel contributes a unified solution tailored to target-label-free and multi-modal fundus analysis. We design a flexible vessel-control strategy that adapts to different modalities. More importantly, we introduce an intermediate modality, i.e., phase maps, to enable the learning and transfer of controllable generation across modalities, providing a new perspective for unpaired controllable generation. The drop mechanism further reduces bias toward public paired data. Together, these designs support multiple downstream tasks without requiring expensive target manual labels. R3-Q2 (the zero-shot claim): We agree that “completely zero-shot” is not rigorous, since UniVessel-img uses target images. Our intended meaning of “zero-shot” is specific to downstream tasks: no real target data are used, and downstream models are trained with UniVessel-generated image-label pairs and evaluated on real target data. We will revise this claim. R4-Q2 (phase map ablation): Canny is not used in UniVessel, but only included to show that our phase map is a better choice of intermediate representation. The DDPM results provide complementary, though not equivalent, evidence for generation without phase maps. Since no new results are allowed, a one-stage no-phase model is left for future work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers are overall positive about UniVessel, converging on three contributions: a physics-informed vessel synthesis framework (L-system + Murray’s law) that produces modality-independent vessel structures with anatomically grounded branching and diameter statistics; a two-step conditional diffusion design (mask → phase-map → image) in which local phase maps serve as an intermediate structural bridge that decouples geometry from device-specific intensity; and an extensive multi-task downstream evaluation spanning zero-shot vessel segmentation (LSCI, PRIME-FP20), retinal image registration (FIRE), and bifurcation landmark detection (FIRE), with gains over baselines.
For the camera-ready, authors should add (i) a no-phase ablation in which $\epsilon_\theta^{MP}$ and $\epsilon_\theta^{PI}$ collapse into a single end-to-end mask→image diffusion, to directly test the claim that phase maps are necessary as an intermediate (the current Canny row is a weak proxy), and (ii) a direct structural-realism evaluation of the synthesized vessel masks — distribution comparisons of vessel density, tortuosity, branching factor, and diameter against target-modality statistics, and/or FID on synthesized images — to substantiate the “modality-aware label generation” pitch beyond downstream-task evidence.