List of Papers Browse by Subject Areas Author List
Abstract
Alzheimer’s disease (AD) is a complicated, heterogeneous neurodegenerative disease associated with cognitive decline, behavioral impairment, and brain atrophy. Detecting individualized pathological changes from cognitive normal (CN) to AD is critical for targeted treatment. Current existing methods face challenges, including biases toward specific pathology profiles. To this end, we proposed a disentangled generative model (DGM) to generate pseudo-healthy images and disease-related residual maps that accurately detect universal pathological changes. The framework of DGM consists of three modules: pseudo-healthy MRI synthesis, residual map synthesis, and input reconstruction modules. We take into account both the healthiness and subject identity to validate the biological validity of synthetic pseudo-healthy images. Our experiments demonstrated the effectiveness of the DGM in reconstructing healthy brain anatomy, preserving subject identity, and highlighting its direct application in anomaly pathological detection across the transitions from CN to MCI and from CN to AD. Code is available at https://github.com/zhonghuajiuzhou12138/DDGM_disease_stage_model.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1508_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{LiZhu_DGM_MICCAI2025,
author = { Li, Zhuangzhuang AND Zhao, Kun AND Wang, Dong AND Liu, Yong},
title = { { DGM: Disentangled Generative Model for Detecting AD Individualized Pathological Changes via Pseudo-Healthy Synthesis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15961},
month = {September},
page = {142 -- 152}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a disentangled generative model for detecting pathological changes in Alzheimer’s disease. By combining a GAN-like approach and a residual map synthesis module, the proposed method generates pseudo-healthy images while generating residual maps from the original input images.
- 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 ultimate goal of the application is clearly defined.
- The attempt to generate normal images from abnormal ones is interesting in itself, and since it is an ill-posed problem, it presents a high level of challenge.
- As intended by the authors, the generated images preserve identity.
- 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 introduction is not properly organized. The background, motivation, and objectives of this study are not sufficiently explained.
- The necessity of generating pseudo-healthy images is not sufficiently explained in the first place. If the goal is to detect pathological changes, then direct classification or segmentation would suffice, and there seems to be no clear reason to solve an ill-posed problem.
- The necessity of the reconstruction module and the residual map synthesis module is not sufficiently explained. Why is it not enough to simply compute the difference between the generated image and the input image?
- There are many errors and unclear points in mathematical expressions:
- In 2.1, $P$ and $H$ are not defined.
- It is unclear whether $x_{p_i}$ and $x_{h_i}$ are sets or elements of a set. While these are written using the symbol $\in$, suggesting that they are elements of a set, the mapping $x_{p_i} \to x_{h_i}$ is defined, implying that they may instead be sets or structures capable of being mapped.
- The distribution $P_X$ is not defined in Eq. (1). It is used as a set in Eq. (2).
- Sets $AD$ and $CN$ are not defined in Eq. (3).
- Please rate the clarity and organization of this paper
Poor
- 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
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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The motivation and objectives are not clearly explained. The necessity of generating pseudo-healthy images using the proposed architecture is unclear. There are many unclear mathematical explanations, making it hard to understand the details of the method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The responses to Q2 and Q3 were not convincing. Regarding Q2, the response does not adequately address my point that direct classification and segmentation would be sufficient. For Q3, I would have liked the authors to state the fundamental necessity rather than just the difference in accuracy.
Review #2
- Please describe the contribution of the paper
Paper propose DGM model, to generate pseudo-healthy images and disease residual map. For doing so they use 3 modules, pseudo-healthy MRI synthesis module (based on generator - discriminator), input reconstruction module and Residual map synthesis module.
- 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 is well written and structured, the system design convincing, and the experimental results are promising.
- 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.
Weakness and Questions: 1) Method defined is very similar to 1. Authors should cite the paper and clearly state the novel contributions as compared to 1. 2) Research problem is not clear to me. Stating research problem - a possible application - numbers to validation will be helpful. 3) Its not clear from the text, which features are used for t-SNE visualization. 4) What is DGM with DRL? It is not clear from the text. 5) Can diffusion model be used for the same task? Something like 2. 6) Can you train a classifier for healthy vs disease classification and check test pseudo-healthy images and report the observation?
Model for Synthesizing Realistic Person Images
- 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.
- 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
I am bit torn, hence
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am bit torn, hence weak reject. Idea is interesting but it is very similar to 1 (see weakness). Paper should clearly state the novel contributions compared to the paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for addressing the queries and appreciate their effort in this submission. I find the problem interesting and the proposed solution appropriate, though the methodological novelty is limited. I would encourage the authors to improve the writing, particularly by adding clearer annotations to the figures and properly citing the related work referenced in the discussion section. Additionally, the comparative analysis could be further strengthened by including more competitive baselines, especially recent diffusion-based approaches.
That said, considering the correctness of the claims made and the potential usefulness of this work for a some of the MICCAI community, I am inclined towards accept.
Review #3
- Please describe the contribution of the paper
The authors propose a disentangled generative model (DGM) to detect individualized pathological changes of Alzheimer’s disease (AD) by generating pseudo-healthy MRI images and disease-related residual maps. The authors decompose pathological images into normal features (shared in health and disease states) and abnormal features (disease specificity), solving the balance problem between healthiness and the retention of the subject identity in pseudo-health synthesis.
- 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 disentangled generative learning was applied to the pathological detection of AD. By effectively disentangling normal and abnormal features, the biological plausibility of pseudo-healthy images was significantly enhanced.
Data validation: EDSD, internal datasets, and ADNI longitudinal data, the biological validity of the generated images was rigorously verified using the longitudinal variations observed in ADNI, thereby strengthening the reliability of the results.
Experimental evaluation: The performance was comprehensively assessed from multiple perspectives, including image quality metrics (e.g., PSNR, SSIM), subject identity consistency (via individual recognition through the SSIM matrix), and pathological correlation (measured by NCC).
- 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.
Insufficient innovation: Disentangled representation learning has similar works in medical images 1, but the essential differences between DGM and these methods have not been clearly explained. 1 Kobayashi K, Hataya R, Kurose Y, et al. Decomposing normal and abnormal features of medical images for content-based image retrieval of glioma imaging[J]. Medical image analysis, 2021, 74: 102227. 2 Yang M, Liu F, Chen Z, et al. Causalvae: Disentangled representation learning via neural structural causal models[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 9593-9602.
Insufficient comparative analysis: The comparative methods (such as CycleGAN and V-GAN) are not specifically designed for medical image untangling and have not fully explained their correlation with the DGM task. The rationality of the comparison needs to be further analyzed. The failure to conduct an in-depth comparison with similar methods in recent times (such as the disease-understanding entanglement adversarial learning in [16]) has weakened the demonstration of the method’s innovativeness.
Insufficient method introduction: The description of the method section is vague, making it difficult for readers to understand the technical details of the method.
- 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
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 — 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 insufficient explanation of methodological innovation and the ambiguity of comparative analysis are the core factors for the weak acceptance of this study.
Firstly, the author proposed a disentangled generative model (DGM). However, they failed to explain the essential differences between DGM and existing similar methods. Secondly, the author conducted experiments to verify the pseudo-image generation quality of the proposed method. However, considering the fuzziness of the experimental analysis and the insufficient depth of biological interpretation of the innovative module, a weak acceptance decision was made.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors addressed my concerns.
Author Feedback
We thank all reviewers and the meta-reviewer for the constructive comments. We have addressed all comments one-by-one below. To reviewer #1: Q1, Q2: Motivation and necessity of pseudo-healthy methods. A: Brain structural changes stand out as predominant hallmarks for Alzheimer’s disease (AD). Modeling AD-related structural changes enables quantification of disease severity, thereby guiding clinical management strategies such as treatment planning. However, it is difficult to obtain longitudinal data to compare structural changes before and after disease onset. Pseudo-healthy synthesis method can solve this problem. Q3: Necessity of the reconstruction and residual map synthesis modules. A: We compared our model with HealthyGAN and DGM without DRL in ablation study, which includes only the pseudo-healthy synthesis module. These methods showed a marked drop in subject identity consistency and slight declines in PSNR, SSIM, and NCC. Q4: Unclear points in mathematical expressions. A: Thanks for your professional comment. We added definitions of the variables to address all concerns. We revised the statement “Let x ∈ P and y ∈ H denote images from the pathological domain and healthy domain; the aim is to learn a mapping a mapping P→H”. To reviewer #3: Q1: Similar to Tang et al. (2021). A: We acknowledge the similarity to Tang et al. and will cite the work. In this paper, we focus on AD-related structural progression changes in 3D MRI, where subtle changes demand more precise localization than visible lesions in X-rays. We modified the network architecture, including the feature extraction module and a patch discriminator, to better detect subtle changes from CN to AD. More importantly, we introduce individual identification accuracy to assess subject identity, since minor generation deviations compromise biological authenticity. Q2: Research problem. A: This study aims to generate pseudo-healthy images and residual maps from disease images to obtain disease-related abnormalities and quantify the severity of AD, thereby guiding clinical management strategies such as treatment planning. Q3: Features for t-SNE. A: t-SNE was applied to normal and abnormal features from the pseudo-healthy synthesis and residual map modules (see Fig. 1). Q4: DGM with DRL. A: DGM with DRL employs disentangled representation learning to build the model, including the pseudo-healthy synthesis, input reconstruction, and residual modules. Q5: Diffusion model for this task. A: In our comparative experiment, a conditional diffusion probabilistic model (cDPM) was used for this task. Q6: Train a classifier for healthy vs disease classification. A: We have completed the classification of the synthesized pseudo-healthy and disease images, but we cannot provide the result due to rebuttal guidelines. To reviewer #4: Q1: Similar to Yang et al. (2021), Kobayashi et al. (2021). A: While Yang et al. achieve identifiability up to permutation, semantic meanings are not guaranteed without strong supervision. Our adversarial design against healthy controls yields explicitly interpretable normal and abnormal factors. Our method shares a similar structure with the network proposed by Kobayashi et al. Building upon their framework, we introduce a discriminator for pseudo-healthy synthesis, along with residual loss and penalty loss to enhance the localization of subtle pathological changes in AD. Q2: Insufficient Comparative analysis. A: Comparative methods (e.g., CycleGAN, VA-GAN) were selected for their representativeness in AD pathology detection but we cannot provide new experiments due to rebuttal guidelines. Future work will explore more disentanglement methods. Q3: Vague description of the method. A: We reorganized the methodology into three parts—Problem Formulation, Disentangled Generative Model (describing the function and network structure of pseudo-healthy MRI synthesis, residual map synthesis, and input reconstruction modules), and Loss Function—to improve clarity.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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”.
Two reviews agree that: The motivation and objectives are not clearly explained. Method is very similar to existing work.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The reviewers tend toward accepting the paper, recognizing the value of applying disentangled generative modeling to pseudo-healthy MRI synthesis for Alzheimer’s disease. The overall architecture is well-motivated, and the experimental validation across multiple datasets strengthens the submission.
While Reviewer 1 raised concerns—particularly regarding the framing of the problem and the necessity of pseudo-healthy synthesis (Q2 and Q3)—these points reflect a preference for alternative formulations rather than fundamental flaws. Clarifying the motivation and refining the exposition would improve the paper, but these issues are not, in themselves, critical grounds for rejection.
Given the positive evaluations from Reviewers 3 and 4, I recommend acceptance, and encourage the authors to expand the discussion of related work—particularly recent diffusion models that have been proposed for UAD and prior pseudo-healthy synthesis methods for AD—as suggested by the reviewers.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A