List of Papers Browse by Subject Areas Author List
Abstract
Accurate modeling of disease progression is essential for comprehending the heterogeneous neuropathologies such as Alzheimer’s Disease (AD). Traditional neuroimaging analysis often confound disease effects with normal aging, complicating the differential diagnosis. Recent advancements in deep learning have catalyzed the development of disentanglement techniques in Autoencoder networks, aiming to segregate longitudinal changes attributable to aging from those due to disease-specific alterations within the latent space. However, existing longitudinal disentanglement methods usually model disease as a single axis factor which ignores the complexity and heterogeneity of Alzheimer’s Disease. In response to this issue, we propose a novel Surface-based Multi-axis Disentanglement framework.This framework posits multiple disease axes within the latent space, enhancing the model’s capacity to encapsulate the multifaceted nature of AD, which includes various disease trajectories. To assign axes to data trajectories without explicit ground truth labels, we implement a longitudinal contrastive loss leveraging self-supervision, thereby refining the separation of disease trajectories. Evaluated on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset (N=1321), our model demonstrates superior performance in delineating between cognitively normal (CN), mild cognitive impairment (MCI), and AD subjects,classification of stable MCI vs converting MCI and Amyloid status, compared to the single-axis model. This is further substantiated through an ablation study on the contrastive loss, underscoring the utility of our multi-axis approach in capturing the complex progression patterns of AD. The code is available at: https://github.com/jianweizhang17/MultiAxisDisentanglement.git
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0002_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jianweizhang17/MultiAxisDisentanglement.git
Link to the Dataset(s)
BibTex
@InProceedings{ZhaJia_Surfacebased_MICCAI2025,
author = { Zhang, Jianwei and Shi, Yonggang},
title = { { Surface-based Multi-Axis Longitudinal Disentanglement Using Contrastive Learning for Alzheimer’s Disease } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {590 -- 599}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed a surface-based multi-axis model that incorporates multiple disease progression axes to more accurately capture the heterogeneity of AD. Additionally, they introduced a novel longitudinal contrastive loss, enabling self-supervised learning to effectively assign data trajectories to their respective axes.
- 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.
This paper has a fancy design of network with contrastive learning, which synchronizes a few losses to update the model prediction. The losses capture information in longitudinal images regarding aging consistency, disease effect and trajectory, while trying to reconstruct the original image.
- 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 authors claimed that they account for heterogeneity of different subjects, but in the results section, I don’t really see it. Subjects are only devided to aging speed and disease speed, and analysis are conducted based on clinical scores and diagnostic stages. No difference in subtype analysis was analyzed.
-
It looked obvious that the aging speed is consistent between diagnostic groups, which is expected. However, the difference in the disease speed is also not significant. So this result is not sensitive enough to detect a group difference.
-
There are five losses in total in the model. Is each loss converging? Is each loss useful? The authors need to do ablation study to show this.
-
In Figure 3, what is p value for here? Is it signifiant? I don’t understand this plot. Why does the color look different for positive and negative values?
-
In Figure 1, since it’s hard to tell physical distribution of errors, it would be nice to decompose a 3d coordinate system plot to a 2D one, and add symbols corresponding to the text. Can squeeze B and add another panel C to illustrate loss functions in this plot.
-
What is the benefit of using a graph neural network rather than a convolutional neural network? Can add a few sentences to explain the choice in the introduction section.
-
- 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
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 results does not persuade me that this is a good longitudinal assessment of disease progression. Nor does it capture heterogeneity of subjects.
- 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
The authors present a framework called Surface-based Multi-axis Disentanglement, aimed at disentangling confounding disease effects from normal aging. The method builds upon a prior autoencoder-based architecture (ref [7]) and extends it by incorporating cortical surface data and multiple disease axes, leveraging a contrastive loss to achieve disentanglement. The results demonstrate promising potential for differentiating among cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer’s disease (AD) groups.
- 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 use of surface-based data for disentanglement is a novel and interesting idea. Surface data may offer more precise cortical representations compared to volumetric data.
-
The introduction of multiple axes for disease modeling is well-motivated, especially considering the heterogeneity and potential subtypes within AD.
-
The experimental evaluation is comprehensive. The inclusion of correlations with clinical assessments adds value and suggests potential clinical applicability.
-
- 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.
-
While the multi-axis modeling for disease progression is well-motivated, the justification for using surface data is less clearly articulated. For instance, key regions affected in AD—such as the hippocampus—are located in the subcortical domain and are not captured by cortical surface modeling.
-
In Section 2.3 (Implementation), more details are needed regarding hyperparameter choices and model sensitivity, especially regarding the weights of different loss components.
-
Certain aspects of the experimental setup require clarification:
- The data partitioning strategy (training/testing splits) is not clearly described.
- In Table 2, it is unclear whether the classifier used is an SVM or a three-layer perception.
- If Table 2 presents 5-fold cross-validation results, it would be helpful to include standard deviations to assess robustness.
-
- 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
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?
My main concern lies in the use of surface data: although this is a novel direction, it inherently excludes subcortical regions that are highly relevant to AD, such as the hippocampus. This omission could limit the model’s comprehensiveness for AD characterization. Nonetheless, the paper presents a technically sound and interesting approach that is worth consideration, particularly due to its innovative multi-axis disentanglement framework and use of contrastive loss in a self-supervision way.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a Surface-based Multi-Axis Disentanglement framework that utilizes contrastive learning to model the heterogeneous progression of AD. By introducing multiple disease axes in the latent space of a graph-convolutional autoencoder, the method disentangles aging effects from disease-specific trajectories, effectively capturing the complexity of AD progression. Additionally, the proposed method incorporates a novel contrastive loss to enhance model 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.
- Introducing multiple disease axes can capture the full spectrum of disease variability, overcoming a significant limitation of previous single-axis approaches. This enables modeling of different disease subtypes and progression trajectories.
- Contrastive learning is employed in a self-supervised manner to construct disease progression, enhancing model performance.
- The strong correlation between age/disease speed and cognitive scores (ADAS, CDR, MMSE) demonstrates clinical interpretability.
- 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 authors define Δz_1，Δz_2，Δz_3 in the early part of Section 2.2, and later introduce Δz_{base}, Δz_{pos}, Δz_{neg}.The relationship and distinction between these sets of definitions remain unclear. Additionally, Equation 7 is confusing as the range of values for j is not specified.
- The contrastive loss defined in the paper operates at the individual level. Could this training approach potentially obscure the learning of the disease axis? Intuitively, the progression direction for same disease should be consistent.
- The constructed data pairs include four pairs. Pair 1 consists of cognitively normal (CN) data from the same subject. Pairs 2, 3, and 4 contain non-CN (MCI, AD) data, with pairs 2 and 3 from the same subject, while pair 4 is from a different subject. Does this imply that all subjects must have data from both the CN and disease stages? In ADNI, data meeting this criterion do not reach a subject size of 1321. Did the authors employ any special training strategies to enable the network to train with subjects containing only disease-stage data or only CN-stage data?
- The authors should provide the subject IDs they used to ensure reproducibility.
- During the training of the classifier, it is not specified whether the rest of the network remains frozen.
- The methods compared in the paper are limited, failing to adequately demonstrate the superiority of the proposed 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?
Overall, the concept of introducing multiple axes for longitudinal disentanglement to model aging and disease progression is novel. However, the method’s description in the article remains unclear, and the comparative experiments are limited.
- 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.
The author has largely addressed my concerns.
Author Feedback
We thank all reviewers for their valuable feedback. Reviewer 1: Q: Lack of Alzheimer’s Disease (AD) subtype analysis A: We use the multi-axis in the latent space to model the heterogeneity and represent diverse progression directions. Compared to the baseline without heterogeneity modeling (single axis), our model achieves better separation of CN (normal), MCI (cognitively impaired), and AD (Table 1), better disease speed correlation with cognitive scores (Fig 3), and better amyloid state classification and cMCI vs. sMCI classification (Table 2). Though we didn’t conduct traditional subtype analysis, we have shown the performance enhancement of adding multi-axis modeling and recognize the potential of our model for explicit subtype modeling. Q: p-value of group difference in disease speed (Fig 2) A: The statistical comparison for Fig. 2 is in Table 1, where the multi-axis (ours) disease speed shows better p-values and Cohen’s d in separation of CN vs. MCI and CN vs. AD than that of single-axis. Q: Convergence and usefulness of each loss A: Our main contribution is the contrast loss for self-supervised AD heterogeneity modeling, and Table 1 contains an ablation study for improved group separation (p-value and Cohen’s d). The first four losses were tested by the previous work [7] through ablation studies and shown to be effective. All losses converge at the end of training. Q: Include loss and better 2D latent space diagram in Fig. 1 (model pipeline) A: To increase clarity, we will include more text and clearer 2D diagrams in Fig. 1 in the final draft. Q: stats and colors in Fig 3 A: Two colors correspond to two disease speeds from our model and the baseline. The Pearson correlations of the disease speed with each cognitive score are shown in the legends, where our model achieves higher correlation than single-axis . Q: Why graph NN? A: We use graph NN because conventional CNNs cannot process cortical thickness data on a 3D mesh, which is our focus for disease modeling. Reviewer #2: Q: Justification for using surface data A: Surface data is more specific to cortical atrophy, which is a widely recognized biomarker for AD. Unlike volume data, it can also capture more intricate atrophy such as cortical thinning. Though our experiment focuses on surface data, our multi-axis framework is general enough to include subcortical volume data in the disease modeling through adding volume encoders. Q: Hyperparameter choices and loss weights A: We want to clarify that we have included network dimensions in Fig. 1 and loss weights at the end of Section 2.3. Q: Confusion on data partitioning, classifier, and lack of std on cross-validation A: For the classification data split, we randomly select, on the subject level, an equal number of cMCI and sMCI for 5-fold CV. The classifier is a 3-layer MLP. We will add the standard deviation of CV in the final draft. Reviewer #3: Q: Definition of Δz_{base}, Δz_{pos}, Δz_{neg} and j in Equation 7 A: The Δz_{base}, Δz_{pos}, Δz_{neg} are Δz2, Δz3, Δz4 (non-CN subjects). (2, 3) are from the same subject for contrast loss. j iterates through Δz2, 3, 4. Q: Individual contrast loss obscures learning common disease axis A: The contrastive loss works on the individual level, but the similarity used to compute contrastive loss is the angular distance from a Δz to each disease axis. Then, data points from the same subject are pulled towards the same axis for consistent disease progression. Therefore, progression direction is in the loss computation. Q: Requirement of CN and non-CN data from the same subject A: We don’t need data with both CN and diseased stages. The CN data defines the aging axis since CN should age without disease effects. The non-CN subjects’ disease components are modeled by subtracting aging components in latent space. Q: Network frozen for classification A: In the classification, the encoders from all methods are frozen, and the latent codes are used for classification.
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”.
N/A
- 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
Meta-review #2
- 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
Meta-review #3
- 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 paper introduces a creative contrastive learning framework with multi-axis modeling for Alzheimer’s disease and offers interesting methodological contributions. However, several important concerns remain insufficiently addressed in the rebuttal, including the lack of explicit subtype analysis, limited ablation evidence for the multiple loss components, and some unclear aspects in the results and figures. Given these remaining gaps, I regretfully recommend rejection at this stage.