List of Papers Browse by Subject Areas Author List
Abstract
Traditional diagnostic groups in Alzheimer’s disease provide only a coarse estimate of disease progression, limiting their utility for early detection and disease monitoring. Recent efforts have sought to learn continuous disease staging from longitudinal neuroimaging biomarkers. However, such data are often limited in practice, as most subjects in typical cohorts have only one single visit. To address this challenge, we propose a self-supervised data augmentation framework based on a residual generative adversarial network (GAN) to synthesize follow-up visits for baseline-only subjects. The model incorporates a monotonic brain change constraint to generate biologically plausible follow-up visits. The synthesized visits showed overall good agreement with real longitudinal follow-up visits in terms of direction of change. When combined with real longitudinal observations, the augmented data lead to consistent and measurable improvements in downstream continuous staging, as evidenced by significantly enhanced temporal preservation across real follow-up visits in held-out test data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0985_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/StochasticAgCl/GAN_SLOPE
Link to the Dataset(s)
BibTex
@InProceedings{LiuHao_Selfsupervised_MICCAI2026,
author = { Liu, Haoyuan AND Tong, Mingzhao AND Fang, Shiaofen AND Yan, Jingwen},
title = { { Self-supervised Longitudinal Neuroimaging Augmentation for Continuous Staging of Alzheimer’s Disease } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The primary contribution of this work is the introduction of SLOPE, a method designed to generate plausible PET images specifically for data augmentation tasks in machine learning applications within neuroimaging.
- 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 key strength of the paper lies in its use of real PET data for training and validation, supported by a well-designed experimental setup that includes a dedicated holdout set.
- 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 of the paper is that the longitudinal constraints embedded in the model introduce a strong bias toward disease progression in the generated images. As the reported results highlight, the generated images exhibit an unrealistic association that is not present in the real holdout data.
- 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 paper could be strengthened by stratifying the results into MCI-AD converters versus MCI-stable subjects. If the proposed data augmentation algorithm improves conversion recall on the holdout dataset, this would provide strong evidence that the approach offers potential benefits for real-world clinical applications.
- 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
This paper addresses the scarcity of longitudinal data in Alzheimer’s Disease (AD) studies by proposing a self-supervised data augmentation framework. The authors use a Residual Generative Adversarial Network (GAN) to synthesize follow-up amyloid PET scans for subjects who only have a single baseline visit. The generator is constrained by a monotonicity assumption (amyloid accumulation is irreversible). The synthesized data is then used to train SLOPE, a continuous staging model. The authors report that this augmentation reduces the “violation ratio” (temporal inconsistency) in the staging scores on a held-out test set.
- 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 problem of sparse longitudinal data in AD cohorts is clinically relevant. 2.The idea of incorporating biological priors (monotonicity) into generative models is interesting.
- 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.Biologically Invalid Assumption of Strict Monotonicity: The core innovation of this paper—enforcing a strict non-negative residual to ensure monotonic amyloid increase—is biologically flawed. While amyloid burden generally increases over the course of AD, it is not strictly monotonic at the individual voxel or regional level 2.Insufficient Validation of Generated Data Fidelity: The paper relies heavily on Fréchet Inception Distance (FID) and cosine similarity of change vectors. FID is known to be insensitive to local structural changes and can be misleading in medical imaging. Cosine similarity measures direction but not magnitude or anatomical plausibility. There is no qualitative assessment (visual inspection by neuroradiologists) provided to confirm that the generated “follow-up” scans look like realistic AD progression. Without demonstrating that the synthetic images preserve anatomical integrity and realistic pathology spread, the claim of “biologically plausible” generation is unsupported. 3.Methodological Ambiguities: The description of the discriminator’s role is unclear. How does the discriminator distinguish between “real” follow-ups and “generated” follow-ups if the generated ones are forced to be monotonic while real ones contain noise? If the discriminator learns to reject noisy real samples in favor of clean synthetic ones, the model may be discarding valuable information about data variability.
- 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.
- 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.
(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?
Limited Novelty and Incremental Contribution: 1.Generative augmentation for medical imaging is a well-explored field. The combination of GANs with monotonic constraints has been explored in other contexts. The application to SLOPE is incremental. Furthermore, the paper does not adequately address why other, more robust methods for handling missing longitudinal data (such as mixed-effects models, multiple imputation, or trajectory modeling techniques specifically designed for sparse data) were not considered or compared. The proposed solution adds significant complexity (training a GAN) for a marginal gain in a specific metric (violation ratio) that may not translate to clinical utility (e.g., prediction of conversion to dementia). 2.The proposed method imposes a rigid biological constraint (strict monotonicity) that does not align with the complexities of neuroimaging data and AD pathology. The resulting improvements in downstream staging appear to be artifacts of this constraint rather than genuine enhancements in disease modeling. Without validation against clinical endpoints or more robust baselines
- 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
The paper describes a generative model for longitudinal amyloid PET data from people with Alzheimer’s disease. The motivation is to use the generative model to augment training data for a staging model. Experiments demonstrate viability of the generator and improvements in a learned staging model after including synthetic data in 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.
-
Well motivated and of high relevance.
-
A simple idea that seems to work reasonably well.
-
Results show a positive impact on the downstream model.
-
- 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 technique (SLOPE) that motivates the work in the submission, and is used in the experiments to demonstrate benefit, it so far only published on biorxiv, so it is yet to be fully verified via peer review. It sounds sensible enough from the description in the paper, but it rather limits the demonstration of the power of the proposed generative model. The synthetic data the model produces would certainly be useful for other models like SLOPE. The paper would be substantially stronger if it showed performance improvements in other methods as well.
-
I found the experiment in section 4.1 a little difficult to appreciate. Unclear what we’re supposed to conclude from figure 2. How does this show that the generative model is doing something sensible when both baselines are flawed?
-
- 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?
Overall this is a reasonable submission. It is not a major contribution, but is a fairly standard data generation method that demonstrably improves training of a potentially useful model.
- 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
Author Feedback
Monotonicity Constraint and Realism of Generated PET Data (R1, R2): We thank the reviewers for raising this important point. We agree that real longitudinal amyloid PET measurements are inherently noisy and may exhibit local or regional fluctuations across visits. Together with the limited number of MCI conversion cases, this presents a major challenge for robust staging of disease progression and may appear to conflict with the monotonicity constraint. In our framework, however, the goal is not to exactly reproduce noisy subject-specific future PET observations. Instead, we aim to generate future amyloid states that are biologically plausible, progression-consistent, and still close to the empirical PET distribution. As a result, the generated samples may deviate slightly from individual noisy observations, which is expected. These cleaner synthetic progression signals are intended to better guide the downstream staging model, helping it learn the underlying biological trajectory of amyloid accumulation rather than overfitting to noises or irregular fluctuations. Additional Downstream Validation (R1, R3): We thank the reviewers for this valuable suggestion. In the current study, we focused on evaluating whether the proposed augmentation improves longitudinal consistency in continuous amyloid staging using SLOPE as the downstream model. We agree that additional validation across downstream tasks and models would provide a more comprehensive assessment of the framework. As the next step of this project, we will extend the evaluation to clinically relevant tasks such as MCI-to-AD conversion prediction and more downstream models beyond SLOPE.
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 are generally positive. It addresses a clinically relevant problem with methodologically interesting and novel approach and is well motivated. While Reviewer #2 mentions weaknesses and reviewer #3 finds the experiment in 4.1 not well enough explained, the work overall has added value to the MICCAI audience