List of Papers Browse by Subject Areas Author List
Abstract
The spatiotemporal changes of a developing anatomical structure is a dynamic process, and quantifying this process within a population and between populations is a fundamental yet challenging task in medical image analysis. Central to this task is the availability of longitudinal imaging data for 4D statistical shape analysis. Unfortunately, this type of longitudinal data is expensive, time-consuming, and difficult to collect. Practically, the majority of imaging data are 3D cross-sectional data, which are inadequate in describing the dynamic shape changes of anatomical structures. In this paper, we introduce a novel temporal atlas-guided deep learning model for longitudinal data generation. Unlike existing methods that directly generate longitudinal data from input images or sequences, we characterize distinctive geometric shape representations in both cross-sectional and longitudinal latent spaces of diffeomorphisms, while optimizing the quality of both atlas and longitudinal data generation. To the best of our knowledge, this is the first deep learning approach that leverages temporal atlas-based representation for longitudinal data generation. The innovative nature of our framework lies in its ability to jointly perform within-age and cross-age shape registration, thus maximizing registration performance while maintaining desirable deformation qualities. Our work’s ability to model spatiotemporal dynamics makes it highly versatile and applicable to a wide range of domains, including modeling the normal and abnormal development of anatomical structures for improved clinical diagnosis and treatment planning. The code of this work is available at https://github.com/wushaoju/TAG-GLE.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1050_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wushaoju/TAG-GLE
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuSha_Temporal_MICCAI2025,
author = { Wu, Shaoju AND Wang, Jian AND Kurugol, Sila AND Tsai, Andy},
title = { { Temporal Atlas-Guided Generation of Longitudinal Data via Geometric Latent Embeddings } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {595 -- 605}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a temporal atlas-guided deep generative model that synthesizes subject-specific longitudinal 3D anatomy (i.e. shape changes over time) from only cross-sectional 3D images. Unlike prior methods that directly predict future images, this approach learns a sequence of age-specific atlas templates and uses them to guide the generation of realistic shape trajectories via diffeomorphic transformations. It is the first deep learning method to leverage a temporal atlas representation for longitudinal data generation, jointly optimizing within-age (atlas) and cross-age (longitudinal) registrations to improve alignment accuracy while preserving smooth, anatomically plausible deformations
- 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 framework combines graph construction with data generation. By learning geometric latent embeddings of diffeomorphic transformations for both cross-sectional and longitudinal spaces, it ensures that the model captures anatomical shape variability across individuals and over time. This temporal atlas-guided approach is novel
- 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.A notable limitation is the narrow scope of data used. The method was demonstrated only on normal pediatric hip CT scans (males age 4–18), with a training set of 82 subjects and a test set of merely 12 subjects for longitudinal evaluation. This is a relatively small and homogeneous sample. It remains unclear how well the model would generalize to other anatomies (e.g. brain, cardiac) or even other populations (female hips, different age ranges) and pathological cases. The current results, while promising, lack validation on diverse or larger datasets, which raises concerns about the robustness of the approach in broader clinical settings. In practice, anatomical variability and disease changes can be much more complex than what’s captured here, so the model might need retraining or adaptation for each new application. 2.Most of the models compared in the paper are older classical models, so I have doubts about the experimental validity of this paper. 3.While the paper claims novelty, it would have been stronger to compare against more learning-based longitudinal prediction methods. The baselines for the generation task were relatively simplistic. This leaves an open question: would a straightforward deep network trained on paired data perform similarly? For example, recent works have begun exploring generative models for longitudinal data, and spatiotemporal atlas learning methods exist. The absence of comparisons to such methods or discussion of why they’re insufficient slightly weakens the argument for originality. In a sense, the paper stakes out a new territory, but the scope of prior art considered is limited. A fair evaluation should acknowledge these emerging methods and clarify how the proposed approach offers advantages.
- 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.
(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?
While the paper presents an innovative framework with promising results, the limited dataset scope, lack of biomechanical or physiological constraints, and issues in figure clarity and explanation reduce its current impact and maturity, warranting a weak reject.
- 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 authors present a deep-learning framework that allows both to estimate spatio-temporal atlases and subject specific longitudinal trajectories. The proposed method leverages geometric latent spaces embeddings. An application is shown with hip CT 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 paper is well written, easy to follow, and properly organized. The presented method makes sense. It is well explained, one can reproduce it from the paper. There are enough novelties to be considered an original method. One can tell that the authors have a good understanding of atlasing and DL-based registration. The choice of the experiments for evaluation is relevant. The results are commented properly. The figures are informative.
- 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.
In the introduction, the authors should talk more about the existing methods and their limitations. The SOTA techniques the authors are comparing against should be briefly introduced. Although quite noticeable with the human eye, a quantitative measure of the sharpness of the produced atlases would be welcome.
- 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
Are phi and A permuted in equation 3? Shouldn’t the arrows in the bottom model of figure 1 point in the opposite direction? Without going into details, it should be indicated in the training setup what kind of DL model Geo-SIC roughly.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The author presented a novel method which makes sense. The paper is clear. There is a proper evaluation. A little downside is that existing SOTA methods are not introduced properly.
- 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 authors have appropriately answered to most of the reviewers comments. The rebuttal answer is quite thorough.
Review #3
- Please describe the contribution of the paper
This paper presents a novel atlas guided, deep learning based, longitudinal 3D image synthesis method incorporating cross-sectional and temporal (diffeomorphic, non-rigid) registration. The method jointly builds a temporal atlas as well as synthesizes longitudinal data. The method generates a sequence of temporal atlases and a model of deformations across ages, which are trained jointly. The trained model synthesizes anatomically plausible trajectories for new images by extracting geometric features through the atlas at the given age. Evaluation is performed on longitudinal CT hip data with large longitudinal deformations.
- 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.
- Methods that capture both cross-sectional and longitudinal trajectories of change are needed and the field is still lacking in good/generally applicable frameworks.
- The joint training of similar embedding spaces for cross-sectional image registration and temporal atlas building is novel
- The evaluation on a presented hip CT dataset (large age range of 4-18 years with significant size and shape changes) is interesting with sufficient benchmarking both of the estimates atlases, as well as the longitudinal synthesis
- Evaluation is performed via segmentation agreement (MSD, dice) of atlas to individual image registration
- The proposed method is novel and of interest to the MICCAI community
- 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.
- Evaluation: no comparison against a method with longitudinal modeling
- Evaluation: rather than segmentations of individual hip joint bone (pelvis, right/left femur), as single region (mask?) is generated and evaluated. Such as segmentation would not be used clinically, also does not providel structure specific evaluation. .
- unclear: Not the full CT image seems to be modeled, but only the region of the hip mask/segmentation. Is that correct? Please clarify in the text and discuss potential limitations.
- modeling is performed in 2x2x2mm imaging space, which is quite low resolution for CT data. In order to relevant, the method needs to be able to deal with higher resolution data.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel method for a timely problem (synthesis of longitudinal data) that is interesting to the MICCAI community. The evaluation is appropriate and indicates that the method performs well for both atlas genertion and longitudinal trajectory synthesis.
- 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 rebuttal is adequate and addressed most of my comments, except for the lack of clarity on the use of a mask vs full image. Still, based on the methodological novelty, I recommend acceptance.
Author Feedback
We sincerely thank the reviewers R1, R2, R4 for their positive feedback and constructive comments, including recognition of our paper as well-written and clear, and our model as novel with strong validation performance. We would like to emphasize that we will release the code and pre-trained models online, accompanied by detailed instructions to ensure reproducibility.
R2 and R4: We thank R2 and R4 for their observations regarding the selection of longitudinal baseline methods. Our approach estimates longitudinal deformations via registration guided by a temporal atlas, which differs fundamentally from learning-based models—such as diffusion or generative networks—that directly synthesize future intensity images. Accordingly, we selected registration-based and atlas-building methods as baselines, as they are more consistent with the core design of our framework. A major benefit of our approach is its ability to generate longitudinal data using only cross-sectional input, without requiring subject-specific longitudinal supervision. This addresses a key challenge in clinical settings, where long-range follow-up imaging is often difficult to obtain. By leveraging atlas guidance, our framework completes the longitudinal growth trajectory in a biologically consistent and data-efficient manner. We will revise the Introduction sections to better highlight this methodological distinction, and we will consider adding learning-based comparisons in future work. R1: We thank R1 the suggestion to expand the literature review and will incorporate a broader overview of relevant methods in the revised manuscript. We also agree that evaluating atlas quality with metrics such as sharpness and clearness is important for assessing image alignment quality, particularly in multi-subject superimposition. We will consider including these metrics in our quantitative evaluation in an extended journal version. R2: We thank R2 for the thoughtful feedback. The decision to focus on the individual hip joint was a deliberate choice for this proof-of-concept study. The hip is a clinically important structure that undergoes substantial anatomical changes during growth, making it an ideal region for demonstrating the capabilities of our method. While this study centers on the hip, our framework is inherently flexible and can be extended to other anatomical regions without modification. We agree that structure-specific evaluations could offer additional insight. We will clarify this in the manuscript and plan to include them in the future. Regarding the 2×2×2 mm voxel resolution, this setting was chosen to reduce training time and facilitate hyperparameter tuning within the constraints of our available GPU memory. We agree that higher-resolution modeling (e.g., 1×1×1 mm) is desirable for clinical applications. Our method is resolution-agnostic and compatible with higher-resolution input, which we intend to explore in future work. R4: We thank R4 for the constructive feedback. The dataset used in this study represents the largest cohort we were able to curate from the imaging archive of a large tertiary hospital, following strict inclusion criteria and a broad search. Although the dataset includes only male pediatric hip CT scans, this well-defined setting enabled us to focus on a clinically meaningful, dynamic anatomical region and establish proof of concept. Importantly, our method addresses the challenge of generating longitudinal trajectories in the absence of subject-specific follow-up scans, a common limitation in real-world datasets. We acknowledge the suggestion to incorporate more recent learning-based longitudinal prediction methods. While our framework follows a fundamentally different modeling paradigm, we agree that discussing emerging learning-based longitudinal methods will add useful context. We will revise the manuscript to clarify how our method complements these efforts. All minor comments will be carefully addressed in the revised version.
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.
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.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A