List of Papers Browse by Subject Areas Author List
Abstract
Reconstructing anatomically faithful vascular structures from clinically accessible imaging modalities is of substantial clinical significance. However, existing cross-modal translation methods mainly emphasize pixel-level fidelity or visual realism and treat structure preservation as a property of the final output rather than an invariant of the generative process. This limitation often leads to structural discontinuities and artifacts, compromising anatomical coherence and clinical reliability. In this work, we propose a Structure-Preserving Mean Flow (SPMF) framework that formulates vascular image translation as a topology-invariant transport process. From a structural invariance principle, we derive an orthogonality constraint on the flow velocity field that formally separates appearance transport from topological distortion, and implement it as a time-weighted surrogate objective within a Brownian-bridge diffusion model to preserve topology at every diffusion step. Moreover, we propose a Prototype-Guided Structural Refinement (PGSR) module to align degraded inference-time structures with reliable training-time structures. Experiments on paired NIRII–2PF and fundus datasets demonstrate consistent improvements over state-of-the-art methods, achieving peak PSNRs of 24.96~dB and 24.83~dB, respectively.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0915_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SunCha_Do_MICCAI2026,
author = { Sun, Changjin AND Hu, Zhuo AND Wang, Kai-Ni AND Wu, Baixuan AND Gao, Shuo AND Zheng, Runan AND Xue, Cheng AND Zhang, Yudong AND Zhou, Guangquan},
title = { { Do Not Break the Vessels: Structure-Preserving Mean Flow for Vascular Image Translation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a Structure-Preserving Mean Flow framework for cross-modal vascular image translation that treats topology preservation as an invariant of the generative process rather than a property of the final output. An orthogonality constraint on the flow velocity field formally separates appearance transport from topological distortion within a Brownian-bridge diffusion model. A Prototype-Guided Structural Refinement module further aligns degraded inference-time structures with training-time prototypes, achieving competitive PSNR results on two vascular imaging datasets.
- 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 addresses an important and clinically relevant problem of vesselness enhancement through cross-modal image translation, with direct implications for improving vascular analysis across imaging modalities.
2 The methodology is well formulated and clearly explains the building blocks of the proposed modules, providing a coherent and technically sound description of the Structure-Preserving Mean Flow framework and its components.
3 The proposed method is properly benchmarked against existing models applied for the same purpose, providing a fair and meaningful comparison that contextualizes the reported performance gains.
- 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 imaging and acquisition protocol is not sufficiently described. Details such as image resolution, anatomical region covered, preprocessing steps, and dataset size should be clearly reported to ensure reproducibility and allow fair comparison with other methods.
2 Figure 3 presents only a single visual example, which is not sufficient to convincingly demonstrate the performance of the proposed method. More diverse examples should be included, and the fundus images are notably absent from the visual assessment, leaving a significant gap in the qualitative evaluation.
3 The paper lacks segmentation and skeleton-wise comparisons after domain translation, which are critical metrics for evaluating structural and topological fidelity in vascular imaging. Relying solely on PSNR without these structure-aware evaluations limits the clinical relevance of the reported results.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
1 The paper addresses an important and clinically relevant problem of vesselness enhancement through cross-modal image translation, with direct implications for improving vascular analysis across imaging modalities.
2 The methodology is well formulated and clearly explains the building blocks of the proposed modules, providing a coherent and technically sound description of the Structure-Preserving Mean Flow framework and its components.
3 The proposed method is properly benchmarked against existing models applied for the same purpose, providing a fair and meaningful comparison that contextualizes the reported performance gains.
- 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 paper proposes a structure-preserving diffusion framework for vascular image translation that treats anatomical topology as an invariant of the generative process, rather than only a property of the final output. Its main contribution is the combination of a trajectory-level structural constraint (SPMF) with prototype-guided structural refinement (PGSR) to improve vascular continuity and reconstruction quality under degraded cross-modal inputs.
- 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 is clearly written and easy to follow. The motivation, method, and experimental sections are presented in a coherent and well-organized manner. 2.The experimental design is reasonable and sufficiently supports the main claims. The comparisons and ablation studies provide useful evidence for the effectiveness of the proposed framework. 3.The work is notably innovative in both formulation and method design. Treating structure preservation as a process-level constraint, together with the prototype-guided refinement module, gives the paper a clear technical novelty.
- 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 evaluation of vascular fidelity is still largely limited to PSNR and SSIM. It would strengthen the paper to include more task-specific structural metrics, such as centerline accuracy, vessel Dice, or connectivity-related measurements. 2.The PGSR module may introduce overly strong prior information from the prototype library. This raises the concern that the model could partially project test samples toward training prototypes rather than faithfully recovering subject-specific vascular structures. 3.Some of the paper’s claims are stronger than what is strictly validated by the implementation. Since the proposed principle is realized through practical surrogate objectives, the wording of the claims should be slightly moderated. 4.The Fundus experiment is somewhat misaligned with the main target task. While it demonstrates robustness under degraded structural observations, it does not fully match the cross-modal vascular translation setting emphasized in the paper.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 submission does not provide sufficient information for reproducibility.
- 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 manuscript presents a generally good work among the manuscripts I reviewed, and I suggest its acceptance.
- 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
This paper addresses vascular image translation, with a focus on recovering anatomically faithful vascular structures when translating from clinically accessible but structurally degraded modalities to higher-quality ones. The primary motivating application is NIRII-to-2PF translation, where NIRII is practical and non-invasive but suffers from scattering and poor spatial resolution, while 2PF provides much clearer vascular detail but is less clinically accessible.
The key observation of the paper is that existing cross-modal image translation methods, including both pixel-wise regression methods and diffusion-based generative models, typically treat structure preservation as a property of the final output rather than as a constraint on the full generative process. As a result, these methods may generate visually plausible images that nonetheless contain anatomically problematic artifacts such as vessel breakage, spurious branches, and topological inconsistencies.
To address this, the paper proposes a Structure-Preserving Mean Flow (SPMF) framework. The main idea is to formulate image translation as a topology-invariant transport process. Starting from the principle that vascular topology should remain invariant along a valid generative trajectory, the paper derives a constraint stating that the flow velocity field should be orthogonal to the structural gradient. Intuitively, this means the generation process should be allowed to change appearance while avoiding motion that would distort the underlying structural manifold.
Since directly enforcing that constraint is impractical with the vesselness operator used in the paper, the authors introduce a surrogate implementation within a Brownian Bridge Diffusion Model (BBDM). At each diffusion step, the model predicts a clean reconstruction from the current noisy sample, and a time-weighted structural consistency loss is applied to encourage the predicted image to preserve vascular topology relative to a structural reference. The weighting emphasizes structural preservation more strongly in earlier diffusion steps, where coarse topology is formed, and relaxes the constraint later to permit finer appearance refinement.
A second contribution is the proposed Prototype-Guided Structural Refinement (PGSR) module. The paper argues that a major difficulty arises from the discrepancy between training and inference: during training, reliable vesselness maps can be extracted from high-quality target images, but at inference time only degraded vesselness estimates are available from NIRII inputs. Directly using these degraded maps as structural guidance is unstable. To address this, PGSR constructs a library of prototype vascular structures from training data, retrieves the most similar prototypes for a given input vesselness map, and refines the input structure by aligning it with these anatomical anchors. The alignment is performed in a frequency-decomposed manner, with greater emphasis on high-frequency components, which are argued to encode topology-critical details such as thin branches and sharp vessel boundaries.
The experimental contribution is an evaluation on two paired datasets: a NIRII–2PF dataset and a fundus dataset. The method is compared against several existing image translation and restoration baselines, including SRGAN, SRN, BBDM, LDL, and SelfRDB. The proposed method achieves the best reported PSNR on both datasets and the best V-MSE, a vesselness-based structural error metric. The ablation study indicates that both SPMF and PGSR contribute meaningfully, and their combination yields the best overall performance.
Overall, the paper’s main contribution is a diffusion-based vascular image translation framework that shifts structural preservation from an output-level objective to a trajectory-level principle, while also introducing a prototype-based mechanism to reduce the train-test structural gap.
- 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) Strong motivation around a clinically meaningful structural problem
The paper is well motivated by the fact that, in vascular imaging, topological fidelity is more important than generic visual realism. Vessel continuity, bifurcation geometry, and the preservation of fine branches are clinically meaningful properties, and the paper correctly identifies that standard image translation methods can fail badly in these respects.
### 2) A conceptually appealing formulation of structure preservation as a process-level property
A major strength is the paper’s attempt to move beyond final-output supervision and instead impose structural consistency throughout the generative trajectory. This is a meaningful conceptual shift. The formulation based on a topology-invariant transport process and the derived orthogonality condition between the structural gradient and the flow velocity provide a clear and elegant interpretation of the problem.
### 3) Integration of structural preservation into diffusion dynamics
The practical implementation within a Brownian-bridge diffusion model is well aligned with the paper’s conceptual goal. The use of a time-weighted structural loss is particularly sensible: stronger topology control in early diffusion stages is a good match to the coarse-to-fine nature of diffusion generation.
### 4) Explicit treatment of the train-test structural gap
The paper identifies an important and realistic issue: high-quality structure is available during training but degraded vesselness is all that is available during inference. This mismatch is often overlooked in conditional medical image generation. The proposed PGSR module is a reasonable and practical mechanism to address that issue.
### 5) Frequency-aware refinement is well motivated
The decomposition into low-, mid-, and high-frequency components, with higher weighting on high-frequency alignment, is intuitive and well justified for vascular imaging, where fine branches and boundaries are indeed structurally important. The paper also provides supporting analysis showing that emphasizing high-frequency components improves vesselness accuracy.
- 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 theoretical claim is stronger than the practical implementation
The paper’s most distinctive claim is the orthogonality constraint on the velocity field derived from structural invariance. However, this constraint is not directly optimized. Instead, the method uses a surrogate structural consistency loss on predicted reconstructions. While reasonable in practice, this creates a gap between the theoretical formulation and what is actually enforced. The paper does not convincingly demonstrate that the surrogate truly approximates structure-preserving flow in the sense claimed by the theory.
### 2) The “mean flow” framing may be somewhat overstated
Although the paper uses the term “Structure-Preserving Mean Flow,” the actual implementation remains a diffusion-based image translation model with an additional structural loss and prototype-guided refinement. It is not entirely clear that this constitutes a genuinely new flow formulation rather than a new regularization strategy applied to an existing BBDM framework. The paper would benefit from more precise positioning relative to prior work on:
- Brownian Bridge Diffusion Models (Li et al., CVPR 2023),
- diffusion guidance and measurement guidance,
- anatomy-compliant or segmentation-guided diffusion in medical imaging,
- structural priors in image translation.
### 3) The structural metric is still a proxy
The paper introduces V-MSE based on Sato vesselness responses, which is more relevant than PSNR alone. However, this remains a proxy metric, not a direct assessment of topology. It may not fully capture clinically meaningful errors such as missed terminal branches, incorrect bifurcations, or false vessel continuity. Additional topology-focused metrics or expert assessment would strengthen the paper.
### 4) Clinical relevance is plausible but not demonstrated
The paper makes a strong case for clinical importance, but there is no direct demonstration of downstream clinical impact, radiologist/clinician evaluation, or use in diagnosis/intervention. Therefore, while the application is clinically motivated, the work remains at the methodological stage.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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 strongest reasons in favor are: 1.the paper addresses a genuinely important problem in vascular image translation, where structural fidelity matters more than generic visual quality; 2.it introduces a conceptually interesting process-level formulation of structure preservation rather than relying only on output-level losses; 3.it proposes a practical mechanism, PGSR, to address the train-test mismatch in available structural information; 4.the empirical results are consistently positive across two datasets and multiple baselines, especially on the structure-related V-MSE metric; 5.the ablation studies support the contribution of both main components.
The main reasons preventing a stronger score are: 1.the theoretical novelty may be somewhat stronger in presentation than in actual implementation, since the key orthogonality constraint is not directly enforced; 2.the “mean flow” framing may not be fully distinct from a diffusion model with additional structural regularization; 3.the prototype-guided refinement, while useful, is not entirely new in spirit relative to prior anatomical prior / guided generation literature; 4.evaluation is solid but still limited, and topology preservation is measured only indirectly through a vesselness-based proxy.
Overall, I find the paper to be meaningful, reasonably novel at the system level, well executed, and likely of interest to the MIC community, but not yet strong enough for a clear accept without reservations.
- 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
N/A
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.
Reviewers agreed that this paper addresses a clinically important problem with a well-motivated and technically sound framework, offering a conceptually interesting process-level formulation for structural preservation and solid empirical results. The main suggestions are to provide clearer dataset/qualitative details and moderate several theoretical claims to better align with the practical implementation. A structure-aware evaluation beyond PSNR/SSIM would further improve the paper quality. Overall, this was viewed positively by reviewers and is considered above the acceptance threshold.