List of Papers Browse by Subject Areas Author List
Abstract
3D blood vessel segmentation remains a critical yet challenging task in medical image analysis. The heterogeneity of clinical imaging protocols introduces substantial domain gaps, limiting the generalizability of supervised learning methods that rely on manually annotated pixel-level labels for individual datasets. Furthermore, the large labeled volumetric datasets are difficult to collect because of data privacy issues. While diffusion models offer potential solutions by generating shareable synthetic data, existing approaches often exhibit poor alignment between synthesized volumes and their corresponding vascular structure input. To address these limitations, we propose Controllable Adversarial Diffusion Model (AVDM), which integrates adversarial supervision into the diffusion training framework. Unlike conventional methods that generate imperceptible perturbations, AVDM synthesizes adversarial instances emphasizing structural variations critical for volume synthesis. Specifically, we design a segmentation-guided discriminator that enforces both the photorealism of generated volumes and pixel-level consistency with original vessel annotations. This supervision mechanism enables high-resolution synthesis of anatomically plausible vascular structures. Experiments demonstrate that AVDM surpasses state-of-the-art methods in generative fidelity and enhances performance on downstream tasks. Our code is available at https://anonymous.4open.science/r/AVDM-C273/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0853_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{DaiJia_AVDM_MICCAI2025,
author = { Dai, Jian and Liu, Wanchen and Cui, Honghao and Liu, Xiao and Wang, Jiajun and Zheng, Zhiji and Geng, Daoying},
title = { { AVDM: Controllable Adversarial Diffusion Model for Vessel-to-Volume Synthesis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {76 -- 85}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper “AVDM: Controllable Adversarial Diffusion Model for Vessel-to-Volume Synthesis” introduces AVDM, a framework that integrates adversarial supervision into a diffusion model to synthesize high-fidelity 3D medical volumes with controllable vascular structures, addressing key challenges in medical imaging such as domain gaps and data scarcity. By combining diffusion models with a segmenter-based discriminator, AVDM ensures anatomical consistency and enables fine-grained control over vessel generation, allowing for applications like vascular disease simulation and synthetic data augmentation. The model achieves high-resolution 3D outputs aligned with real vessel annotations while demonstrating superior cross-domain generalizability. Key contributions include a hybrid adversarial-diffusion architecture for improved fidelity, controllable synthesis via conditioning on vessel structures, and robust performance in downstream tasks like segmentation. AVDM’s ability to generate realistic, customizable volumes has significant implications for medical AI, reducing reliance on scarce annotated datasets and enhancing diagnostic model training.
- 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.
- Good theoretical framework: Using recent advanced models to target a challenging problem, Consideration of the problems of domain shift and the lack of access to bulk annotated data in medical applications
- 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.
- Lack of Originality: This study is based mostly on existing models. These models include references [13] and [14]. Similar examples to the proposed model have already been published such as (https://doi.org/10.1007/978-3-031-72933-1_6) and (https://doi.org/10.1007/978-3-031-72104-5_9). Besides, in the last paragraph of page 4, the sentence “Inspired by, we formulate…” does not mention which reference the section in question was inspired by.
- Methodological Flaws: The results of the ablation studies in Table 4 do not seem to be able to lead to a precise choice of the appropriate model, because the number of parameters, the architecture of the models used and other issues can overshadow the choices. Besides, the results are only compared with well-known generative methods and there is no suitable comparison with related works presented in the same field.
- Poor Flow: The article is hard to follow and lacks proper coherence in the text. For example, the abstract does not adequately describe the research conducted, its objectives, and the study’s contributions. The study’s contributions are not explicitly mentioned in the introduction, and due to the lack of a proper literature review, the challenges of the subject and their relationship to the study’s contributions are unclear.
- Superficial Literature Review: The relevant studies have not been well reviewed and their challenges have not been addressed.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/en/REVIEWER-GUIDELINES.html
- Relevant studies can be presented in a paragraph in the introduction, and their challenges can be mentioned.
- Appropriate comparisons with these relevant studies can be presented to better understand the current study’s position among related studies. The abstract should also be rewritten so that the aims and contributions of the study can be understood. The abstract is a bit confusing in its current form.
- AVDM stands for “Controllable Adversarial Diffusion Model” which doesn’t seem like a good abbreviation. It also doesn’t match the abbreviation mentioned on the authors’ GitHub.
- The formula on page 4, second paragraph, should be corrected.
- I recommend that the authors mention potential future work in a single sentence.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The lack of originality, methodological flaws, poor flow of the paper, and superficial literature review made me make this decision. Despite the lack of originality, the presented model is based on advanced and recent methods to target this challenging problem.
- 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 #2
- Please describe the contribution of the paper
This paper introduces a novel generative model, AVDM (Controllable Adversarial Diffusion Model), for 3D vascular segmentation. The main contributions of the work can be summarized as follows:
A. The authors address a key limitation of existing diffusion models—namely, the misalignment between the synthesized volume and the input vascular structure—by integrating adversarial supervision into the diffusion training framework.
B. They propose a method for generating structurally faithful and domain-adaptive adversarial examples by traversing along optimized gradient directions in a low-dimensional latent space.
C. A segmentation-guided discriminator is designed to jointly enhance the realism of the generated volumes and ensure pixel-level consistency with ground truth vascular annotations.
- 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.
A. Clarity of Problem Definition: The authors clearly articulate the significance of 3D vascular segmentation and the limitations of current methodologies. In particular, they effectively highlight the domain gap caused by heterogeneity in clinical imaging protocols, as well as the challenges associated with manual annotation.
B. Methodological Approach: The integration of adversarial supervision into the diffusion model framework is an intriguing strategy. Notably, the use of a segmentation-guided discriminator to enhance the structural consistency of the generated images is a commendable contribution.
C. Experimental Design: The authors conduct a comprehensive evaluation of their proposed method across diverse datasets (ADAM, CoW, IXI-HH). They systematically assess its performance on downstream tasks such as zero-shot, one-shot, and few-shot segmentation, providing strong empirical validation.
D. Quantitative Results: The proposed model outperforms state-of-the-art methods in FID and MMD metrics, and demonstrates superior performance in downstream segmentation tasks as well.
- 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.
A. Lack of Theoretical Justification: The theoretical rationale for integrating adversarial supervision into diffusion models is not sufficiently convincing. The manuscript lacks a thorough theoretical analysis explaining why this approach is expected to outperform existing methods. In contrast, Yang et al. (2022) provide an in-depth theoretical foundation and broad analysis of diffusion models across various applications, which is currently missing in the justification for AVDM.
B. Limited Choice of Baselines: There is some concern regarding the selection of comparative methods (HA-GAN, WGAN, WDM, MAISI), as it is unclear whether they represent the current state of the art. In particular, the paper lacks comparisons with more recent diffusion-based generative models for medical imaging. Cao et al. (2022) offer a comprehensive comparative analysis of recent diffusion models and their applications, which should be considered to strengthen the experimental evaluation.
C. Lack of Clarity in Method Description: The methodology section is heavily reliant on mathematical formulations, which makes it difficult to intuitively grasp the underlying mechanisms. Specifically, the meaning and implementation details of Equations (3) and (7) are insufficiently explained. Croitoru et al. (2022) demonstrate effective strategies for intuitively presenting the mechanics of diffusion models, and a similar level of clarity would be beneficial for this work.
References: Yang, L., et al. (2022). Diffusion Models: A Comprehensive Survey of Methods and Applications. arXiv:2209.00796 Cao, H., et al. (2022). A Survey on Generative Diffusion Model. arXiv:2209.02646 Croitoru, F. A., et al. (2022). Diffusion Models in Vision: A Survey. arXiv:2209.04747
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/en/REVIEWER-GUIDELINES.html
While the authors contribute to the development of generative models for 3D vascular segmentation, several areas require improvement:
A. A clearer and more intuitive explanation of the proposed methodology is needed. In particular, the integration of adversarial supervision into the diffusion model and its advantages should be elaborated in greater detail to enhance understanding.
B. Where possible, an evaluation of the clinical utility of the generated images would strengthen the work. This could include a qualitative assessment by medical professionals or a discussion on the potential applicability of the generated outputs in real-world clinical settings.
C. A more comprehensive comparison with recent diffusion-based generative models for medical imaging appears necessary to better position the proposed method within the current state of the art.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors make a meaningful contribution to the development of generative models for 3D vascular segmentation. The experimental results are compelling, consistently outperforming state-of-the-art methods across multiple metrics including FID, MMD, Dice, and clDice.
A. Strength of Experimental Results: The proposed method demonstrates superior performance across a variety of metrics. In particular, the results presented in Table 1 and Table 2 clearly show that the method significantly outperforms existing approaches.
B. Importance of the Addressed Problem: The problem targeted by this work—generating high-quality synthetic data for 3D vascular segmentation—is highly relevant and important in the field of medical image analysis.
C. Methodological Integration: Although the approach is not entirely novel, the effective integration of diffusion models with adversarial learning is noteworthy and represents a well-executed methodological contribution.
- Reviewer confidence
Somewhat confident (2)
- [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 introduces AVDM, a novel diffusion-based framework designed for generating high-fidelity volumetric medical images conditioned on vessel masks. By incorporating a diffusion inversion mechanism, the method maps input images into a latent space and reconstructs them while preserving anatomical structures and textures. The use of a semantic segmentation-based discriminator provides fine-grained feedback during generation, supporting structural fidelity. Overall, the proposed method demonstrates promising results and appears to be a meaningful advancement in mask-conditional image generation.
- 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 dataset setup is a strong point,; evaluation conducted on three external datasets, which supports generalizability.
- The model design is thoughtfully constructed, leveraging semantic feedback and latent space manipulations to enhance image realism and structural accuracy.
- Included ablation studies that demonstrate the contribution of each component in the architecture.
- 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 manuscript mentions the extraction of 50 patches per patient to ensure vessel representation, but it does not explain the strategy or criteria for patch selection. Clarifying whether the selection is random, vessel-density-driven, or spatially stratified would strengthen reproducibility and help assess potential biases in training. -The use of diffusion inversion (projecting an image into the latent space at a specific timestep) is an interesting twist on the denoising process. However, this design choice should be more clearly justified. Has this approach been used in prior work? If so, please cite it; if not, explain the rationale and advantages of this inversion method compared to standard approaches. -The paper would benefit from a dedicated section discussing current limitations (e.g., dependence on vessel masks, inference speed, generalization to other organs/modalities) and outlining potential future directions
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Please correct the typos in:
- “In contrast, combining Latent Optimization and ControalNet achieves the best segmentation results.” → ControlNet -Table 3: Correct “componen” → component
- 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 — could be 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 an interesting approach to vessel-conditioned volumetric image generation using diffusion models, with a strong dataset setup and thorough ablation studies. While some methodological details (e.g., patch selection and diffusion inversion) need clarification, the paper lacks a discussion of limitations or future work.
- Reviewer confidence
Somewhat confident (2)
- [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 sincerely thank the reviewers for their time and effort in thoroughly reviewing our manuscript. We greatly appreciate their recognition of its contributions and their insightful, constructive suggestions, which have significantly strengthened our work and provided valuable guidance for our future research. Reviewer #1: To address the missing reference in the sentence “Inspired by, we formulate…” on page 4, we apologize for the oversight. we will explicitly cite in the revised manuscript. We acknowledge that the ablation studies in Table 4 may lack precision in guiding model selection due to variations in model parameters and architectures. We will provide detailed information on all the ablated configurations,. Reviewer #2: We thank the reviewer for noting the need for a more robust baseline selection. Our original baselines (HA-GAN, WGAN, WDM, MAISI) were chosen for their established use in medical imaging, representing diverse generative approaches (GAN-based and diffusion models). We will compare with more diffusion-based generative models for medical imaging in future work. Reviewer #3: We employ a vessel-density-driven approach, where patches are selected based on the presence of vascular structures identified in ground-truth vessel masks. A sliding window algorithm scans the 3D volume, prioritizing regions with high vessel density (computed as the proportion of vessel voxels within a patch). For each patient, we select the 50 patches ranked by vessel density, ensuring at least 20% vessel coverage per patch to balance representation and diversity. We thank the reviewer for suggesting the inclusion of a limitations and future directions section. We will optimize inference speed through techniques like distilled diffusion models and evaluate AVDM on diverse datasets to assess cross-modal generalization in future work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A