List of Papers Browse by Subject Areas Author List
Abstract
Accurate prediction of disease progression from longitudinal medical imaging is critical for personalized treatment planning, yet remains challenging due to subtle progression cues, scan-to-scan variability, and uneven visit timing. Existing methods often model either temporal consistency across visits or temporal change signals, but rarely capture both in a unified progression framework. We propose TeDyS, a temporal dynamics–conditioned query framework that jointly models short-term and long-term progression patterns under variable inter-visit spacing. TeDyS maintains a set of temporal queries that are conditioned on explicitly decomposed temporal dynamics, including short-term aligned residual changes and long-term baseline-referenced trends. This design enables robust modeling of both fine-grained progression signals and long-term disease trends within a compact latent representation.
We evaluate the proposed framework on an in-house longitudinal pancreatic CT cohort and a public longitudinal brain MRI dataset, assessing its performance across different organs and imaging modalities under irregular real-world follow-up. Code will be released at https://github.com/Ewha-AI/TeDyS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4449_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Ewha-AI/TeDyS
Link to the Dataset(s)
OASIS2 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-2/
BibTex
@InProceedings{HwaJeo_TeDyS_MICCAI2026,
author = { Hwang, Jeonghyun AND Roslan, Nur Syamiela AND Jeong, Jisoo AND Woo, Sang Myung AND Viriyasaranon, Thanaporn AND Choi, Jang-Hwan},
title = { { TeDyS: Temporal Dynamics with Spatial Context for Longitudinal Progression Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This study focuses on addressing the challenge of modeling disease progression on longitudinal imaging accounting for variations in follow up intervals, batch effects, subtle changes. They present TeDys which takes into account short term and long term dynamics in disease progression. They evaluate TeDys on pancreatic CT and brain MRI cohorts with longtiduinal imaging. They report ablation studies and comparisons across competing methods demonstrating consistently superior 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.
1) Modeling variations in follow-up in longitudinal imaging datasets is clinically relevant. Modeling in terms of long term and short term residuals with respect to follow-up time from baseline is interesting. 2) The introduction is very well written providing context for including temporal variations in longitudinal imaging based disease progression modelling. 3) Evaluating on 2 different case studies is a strength.
- 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) The modeling seems similar to a typical LSTM architecture which has been studied quite a bit (example PMID: 38341100). While the authors compare TeDys to LSTM, they haven’t really mentioned LSTM anywhere in the introduction or discussion. It is not very clear how this is unique and different compared to existing methods. 2) Feature encoders are directly used from pre-trained segmentation models. Can the authors compare with other foundation models such as MedSAM or SAMMed3D? Is there any rationale for using nnUnet? 3) Mean pooling of short and long term seems too simplistic. How are the long term and short term dynamics interacting with each other? 4) It is unclear what is the attention based on? Do the authors look at disease specific regions of interest form the scans to track changes? Perhaps more detail about the regions of interest or some qualitative visualizations would have helped understand the disease progression better.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the problem was motivated well, the experiments and evaluation fell short. They use very specific encoders without providing rationale. They don’t really demonstrate how the limitations of previous methods are being addressed. Dataset descriptions of region of interest are missing.
- Reviewer confidence
Very confident (4)
- [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 rebuttal clarifies experimental details, statistical robustness. However, I did not get additional clarity on the novelty over existing works which also look at longitudinal progression data using LSTM that has a very similar approach with different semantics. Details on biological significance of attention was also not addressed.
Review #2
- Please describe the contribution of the paper
The authors address the task to base conversion predictions for diseases on longitudinal scan sequences instead of singular scans. They explicitly decompose the short-term dynamics (STD) and long-term dynamics (LTD) of the disease trajectory using an optimal transport application and a neural ODE block in their architecture.
- 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 main novelty is the explicit treatment of long- and short-term dynamics in longitudinal prediction tasks using appropriate techniques. This is achieved by explicitly splitting the long- and short-term dynamics into two different mechanisms: entropic optimal transport for the short-term and a neural ODE block for the long-term dynamics.
- 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 proposed method demonstrates improvements on the private dataset but yields performance comparable to existing baselines on the public dataset. Given the large standard deviations reported, it is unclear if the observed improvements are statistically significant or robust across different cohorts.
It is unclear if the 5-fold cross-validation was stratified by progression intervals (the target labels). Given the small sample size and high variance in metrics, a lack of rigorous stratification raises concerns about potential data leakage or spurious findings.
The utility of the public dataset for evaluating longitudinal progression is questionable, as most subjects have only two visits. Furthermore, with only 14 conversion cases (~9%), a 5-fold split leaves very few converter samples per fold. There is a risk that the model is performing static disease classification rather than capturing true prognostic trajectories. If this is intended, it should be mentioned as a limitation.
The ablation studies suggest that the decomposed STD mechanism does not substantially improve performance over the LTD block. This calls into question the practical necessity of the added architectural complexity.
- 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.
- 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
Overall clarity of the architecture: It would be beneficial if the authors could provide additional information on the specifics of some parts of the architecture choices, as some operators and variables are not defined in the text. Although they might be inferred from the provided source code, it would be helpful to have a full description in the paper itself.
Methodological Rationale: It would be beneficial if the authors could provide further insight or a more detailed justification for the selection of entropic optimal transport for capturing short-term dynamics.
Experiments: To further demonstrate the model’s robustness, the authors are encouraged to evaluate the approach on a larger longitudinal dataset (such as OASIS3). Additionally, please clarify if stratification was performed at both the patient and conversion-label levels. If this is not possible, add a confusion matrix to make transparent which labels were identified correctly (and if they are actual converters).
Ablation Studies: Performing additional ablation studies on a larger dataset would help to more clearly quantify the specific benefits of the proposed short- and long-term decomposition.
- 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The small sample size of the validation and unclear stratification of the 5-fold CV makes it difficult to judge the true improvement of the proposed method and the generalizability.
- 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 TeDyS, a temporal dynamics-conditioned query framework for longitudinal disease progression prediction from multi-visit imaging. The method explicitly decomposes temporal information into short-term residual changes and long-term baseline-referenced trends, and integrates them via a query-based representation. Experiments on an in-house pancreatic CT cohort and the public OASIS-2 brain MRI dataset show consistent improvements in AUC and recall under irregular follow-up.
- 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 paper addresses a clinically important and realistic problem of longitudinal progression prediction under irregular follow-up, with clear relevance to patient management and treatment planning. 2.The proposed framework introduces an explicit and well-motivated decomposition of temporal dynamics into short-term and long-term components, providing a more structured alternative to standard sequence modeling approaches. 3.Experimental validation is solid, including evaluation across two datasets and imaging modalities, with consistent improvements over longitudinal baselines and supporting ablation studies.
- 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.While the overall framework is coherent, many individual components (e.g., query-based modeling, Neural ODEs, optimal transport alignment) build on existing techniques. The contribution is primarily an integration rather than a fundamentally new modeling paradigm. 2.Although ablations are included, the framework remains relatively complex, and further analysis isolating individual components would improve interpretability. 3.The in-house dataset is relatively small, and although a public dataset is included, validation across additional cohorts or institutions would strengthen claims of generalizability.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(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 well-motivated and technically sound framework for longitudinal progression modeling, with a clear formulation and consistent empirical gains across datasets. While the methodological novelty is moderate and the model is somewhat complex, the structured treatment of temporal dynamics and solid validation support acceptance.
- 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 stay with my original decision and reasoning.
Author Feedback
We sincerely thank the reviewers and meta-reviewer for their thoughtful feedback. We address the main concerns below.
[R2, R4W3: Dataset and statistical reliability] All experiments used patient-level 5-fold CV, with all longitudinal visits from the same subject kept within the same fold, preventing scan-level leakage. Fold construction considered progression-interval/time-to-progression information to preserve progression-related heterogeneity across folds. We further audited the binary conversion/progression distribution and confirmed that every fold contained positive cases. We will report fold-wise label counts and converter-level confusion matrices in the revision. OASIS2 was included to evaluate whether the trend generalizes to public brain MRI, while its limited converters/visits may contribute to fold-wise variance. Within this evaluation, conversion cases were represented across folds, and TeDyS achieved approximately 71% accuracy on these cases. In the larger ADNI MRI cohort, TeDyS achieved the best numerical performance, improving accuracy (0.7887→0.8098) and AUC (0.8501→0.8606) over the strongest baseline, while better identifying positive cases in the binary task (0.6379→0.7586; McNemar’s mid-p p=0.0287; bootstrap p=0.0274). We will clarify the OASIS2 limitation and discuss larger-cohort validation.
[R1W1, R4W1: Difference from LSTM and novelty] Unlike LSTM variants that implicitly accumulate temporal evidence in recurrent hidden and cell states, h_t and c_t, TeDyS explicitly constructs two progression signals before query updating: Delta_STD_t = X_t minus X_tilde_(t-1 to t) for aligned adjacent-visit change, and Delta_LTD_t = X_t minus X_hat_(t given 1) for baseline-referenced long-term deviation. These residual dynamics directly condition the temporal query updates, while the TeDyS gate adaptively balances STD and LTD evidence at each visit. Thus, TeDyS is not a generic recurrent memory model or a simple combination of existing modules, but a progression-specific temporal decomposition framework that separates, aligns, and integrates complementary short- and long-term evidence for longitudinal prediction.
[R1W3, R2W4, R2C1C2, R4W2: Necessity of STD/LTD complexity] We agree that LTD dominates overall AUC, while STD targets short sequences where long-term trajectories are weak. In-house, STD outperformed LTD for two-timepoint cases (AUC 0.5872 vs. 0.5370), whereas LTD was stronger for >=3 visits (0.8001 vs. 0.5389), supporting their complementary roles across sequence lengths. Although the AUC gain from adding STD is modest, the full model improves Acc/Recall/F1 over LTD-only, suggesting improved progression sensitivity and short-sequence robustness. Entropic OT is used to align slot tokens across visits before computing STD residuals, since slot ordering is not temporally fixed. Mean pooling only summarizes STD/LTD evidence for the gate, while token-level residuals are fused before query update. The added complexity is modest: STD-only, LTD-only, and STD+LTD use 4.29M, 4.66M, and 5.38M parameters, respectively. We will also add brief qualitative progression visualizations showing that the temporal attention highlights regions with longitudinal changes.
[R1W4: Encoder selection] We selected nnUNet as a practical medical-image backbone because its segmentation-oriented features provide organ/lesion-aware spatial representations relevant to disease progression, with manageable computational cost. TeDyS is not restricted to nnUNet, as the same temporal modeling pipeline can be applied to MedSAM, SAMMed3D, or other foundation encoders.
Overall, TeDyS models short-term visit-to-visit changes and long-term baseline-anchored trends as complementary temporal evidence for irregular longitudinal imaging. The in-house CT, public MRI, and larger-cohort results support TeDyS as a structured alternative to implicit recurrent temporal aggregation across different modalities, anatomies, and disease settings.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
The paper presents an approach to model longitudinal disease progression by explicitly decomposing temporal dynamics into short-term and long-term components. While the use of Neural ODEs is technically sound, the authors must address significant concerns regarding the method’s unique contribution compared to standard LSTM architectures. Reviewers noted a critical lack of detail concerning the absence of clear stratification in the cross-validation process and the potential for data leakage given the small N. Additionally, the modest performance gains on public datasets, coupled with high standard deviations, raise questions about the statistical significance of the findings. The authors should also justify the necessity of the added architectural complexity, as current ablation studies suggest the short-term dynamics module may not substantially improve performance. Finally (probably not super critical), providing a clear rationale for the chosen feature encoders and qualitative visualizations of the disease-specific regions of interest would greatly strengthen the paper.
- 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.
Concerns about methodological rationale and advantages over LSTM-based methods are not adequately addressed post-rebuttal.
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.
While there are some reservations on the novelty, the framework is coherent and the method was evaluated on 2 different case studies, which were highlighted as strong points. Therefore the paper is leaning towards accept.
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.
The paper presents an interesting methodology for a critical problem relevant to the MICCAI community. Even though the two datasets are relatively small, the experimental setup is well-designed with adequate baselines and clear improvements.