Abstract

Ischemic stroke is a major global disease. Treatment decisions are highly time-sensitive, as eligibility for reperfusion therapies relies on the interval between stroke onset and intervention. However, the true onset time is often uncertain in clinical practice, necessitating imaging-based assessment of tissue age as a surrogate marker. Early ischemic changes on routinely acquired non-contrast CT (NCCT) are often subtle, and real-world clinical datasets exhibit pronounced onset-time class imbalance and center-scanner-related heterogeneity. In this work, we propose StrokeTimer, a fully automated framework for onset-time estimation in acute ischemic stroke. StrokeTimer integrates self-supervised disentanglement learning with energy-guided contrastive learning to capture subtle ischemic patterns while addressing long-tailed data distributions under acquisition variability. Onset time is categorized into three clinically relevant windows (<4.5 h, 4.5–6 h, and >6 h). Experimental results on a large multi-center NCCT dataset from two national cohorts, show that StrokeTimer achieves a macro AUC of 0.69 and a macro F1-score of 0.57, improving the strongest baseline by nearly 50% (p < 0.005). In this realistic, challenging setting, representative baseline approaches exhibit near-chance macro performance. Model explanations further highlight subtle gray–white matter blurring and hypodense regions consistent with established radiological biomarkers. These findings demonstrate the potential of StrokeTimer to support treatment decision-making in acute ischemic stroke. We will make the code publicly available.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0549_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/BrainVas/StrokeTimer.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanWei_StrokeTimer_MICCAI2026,
        author = { Wang, Weiru AND Olthuis, Susanne G. H. AND Lavrova, Elizaveta AND van Oostenbrugge, Robert J. AND Majoie, Charles B. L. M. AND van Zwam, Wim H. AND Su, Ruisheng},
        title = { { StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-Contrast CT } },
        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

    They propose a method to predict stroke onset time across three windows (<4.5 hours, 4.5–6 hours, and >6 hours), which is a critical factor for determining appropriate therapy in ischemic stroke. The approach uses a two-stage framework that does not require manual lesion annotations. The first stage performs self-supervised semantic–style disentanglement, while the second stage applies an Energy-Guided Contrastive Mean-Shift (ECMS) method. The authors also compare their approach with other self-supervised learning (SSL) methods in a long-tail classification setting.

  • 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 combination of semantic–style disentanglement and the ECMS module is an interesting and novel approach.
    • Not requiring lesion annotations is a strong advantage, as it reduces the need for costly and time-consuming manual labeling.
    • The use of an NCCT dataset from 18 medical centers strengthens the study by demonstrating diversity and potential robustness across different sources.
    • The use of Grad-CAM++ heatmaps for visualizing stroke onset is valuable for interpretability, although the visualizations and their explanations (e. g. , relationships between normal vs. abnormal regions, onset time, and slice selection) could be clearer.
    • The proposed StrokeTimer shows improved performance compared to other SSL methods, although the overall performance remains relatively modest
  • 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.
    • Although StrokeTimer improves over other methods in long-tailed classification (both per class and on average), the overall performance remains relatively low, mostly in the 0.5–0.7 range.
    • In Table 2, some methods achieve REC= 1.00 for the <4.5 class but are not highlighted, which raises concerns about result presentation and consistency (please revise this). Overall, the performance gains do not appear strongly significant across all classes.
    • The method is relatively complex due to the use of self-supervised learning, yet the improvement over non-SSL approaches seems limited, which may not fully justify the added complexity.
    • The evaluation is limited to a single dataset; testing on additional datasets in the future (for journal submission, for example) would help assess the robustness and generalizability of StrokeTimer.
  • 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.

    (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 proposes an interesting method for predicting stroke onset time using NCCT, while the current gold standard for ischemic stroke assessment is MRI/perfusion imaging. The overall aim and methodology are strong. However, some improvements are needed, particularly in the presentation of tables and in making the Grad-CAM visualizations clearer and easier to interpret.

  • 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

    This paper proposes a novel Ischemic Stroke Onset-Time Estimation framework for detecting Ischemic Stroke (IS) lesions. Overall, the manuscript is well organized and easy to follow. The extensive experiments conducted on two national cohorts (1886 NCCT scans) are particularly impressive, demonstrating the effectiveness of the proposed method and clear improvements over competing approaches. In addition, the gram-based visualization analysis provides valuable insights and enhances the interpretability of the method. The work is clinically meaningful, especially considering the advantages of NCCT in terms of cost and acquisition speed, which could potentially reduce the burden of stroke diagnosis and treatment.

  • 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.

    Traditionally, the estimation of stroke onset time has relied primarily on multimodal MRI. The authors have innovatively utilized non-contrast CT (NCCT) to estimate onset time, proposing a novel methodology specifically tailored to the characteristics of the data. This approach holds the potential to enhance the efficiency of stroke diagnosis and assessment while simultaneously reducing examination costs.

  • 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 clinical motivation of this work is highly appreciated. Traditionally, stroke onset-time estimation has relied on multi-modality MRI, where infarct characteristics can be better distinguished across different imaging sequences. Moreover, prior studies often depend on lesion ROI-based feature extraction. Given that early ischemic changes on NCCT are typically subtle and difficult to perceive, it would be helpful if the authors could further clarify how the proposed method is able to effectively focus on lesion-related regions without explicit lesion segmentation. Please specify it. 2.In the first component of the method, the authors introduce a Semantic-Style Disentanglement strategy by leveraging frequency-based encoders to extract semantic and style representations. It would be beneficial to further elaborate on how these disentangled representations specifically contribute to the downstream stroke onset-time estimation task. Providing more intuition or empirical evidence could improve the clarity of this design. 3.Furthermore, it would be helpful if the authors could clarify how the learned disentangled representations are transferred to the downstream task. In particular, more details regarding the fine-tuning strategy or integration mechanism would improve reproducibility and understanding. 4.For Fig. 2, it is suggested to plot the infarct regions (e.g., annotated on skull-stripped NCCT images), so that the focus areas of different methods’ heatmaps can be more clearly interpreted and compared. Overall, this is a well-motivated and promising work with strong experimental validation. Addressing the above points would further improve the clarity and impact of the paper.

  • 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.

    (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?

    I have previously conducted in-depth research related to the subject of this paper and have discussed the potential clinical value of undertaking this study with neurosurgeons. Therefore, I believe that this work possesses significant research value.

  • 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 #3

  • Please describe the contribution of the paper

    This paper proposes StrokeTimer, a fully automated framework for estimating ischemic stroke onset time windows from non-contrast CT (NCCT). The key contributions are: (1) a self-supervised semantic‑style disentanglement module to handle acquisition‑related variability; (2) an Energy‑guided Contrastive Mean‑Shift (ECMS) module to address long‑tailed class distributions; and (3) the first fully automated benchmark for NCCT‑based onset estimation without manual lesion delineation. Experiments on a large multi‑center dataset (1,686 scans from 18 centers) show that StrokeTimer achieves a macro AUC of 0.69 and macro F1 of 0.57, significantly outperforming several long‑tailed and self‑supervised baselines.

  • 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.

    Estimating stroke onset time from routine NCCT is an important and challenging task, especially when the true onset is unknown. The paper tackles a real‑world bottleneck in acute stroke triage.  The framework explicitly models both imaging style variability (scanner/center differences) and severe class imbalance (only 72 samples in the 4.5–6 h window). This integrated treatment is a genuine strength.  The dataset is large (1,686 scans) and multi‑center (18 sites), which is a major step beyond single‑center studies. The ablation studies clearly demonstrate the contribution of each component (disentanglement vs. ECMS).  Grad‑CAM++ visualizations show that StrokeTimer’s attention maps focus on clinically meaningful regions (gray‑white matter blurring, hypodensity), enhancing trustworthiness.

  • 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.Limited novelty in methodological components. The semantic‑style disentanglement via VAE + FiLM and the prototype‑based contrastive loss are well‑established techniques in the medical imaging literature (e.g., Sarhan et al. ECCV 2020 for orthogonal disentanglement; Li et al. ICLR 2021 for prototypical contrastive learning). The ECMS module essentially combines a weighted mean‑shift update with a frequency‑based prior – a heuristic that lacks theoretical justification.  2.Insufficient comparison with recent task‑specific methods. The paper only compares against generic long‑tailed classification baselines (Focal Loss, Balanced Softmax, etc.) and self‑supervised methods (SimCLR, MoCo). However, there exist several recent deep learning approaches specifically designed for stroke onset estimation from NCCT (e.g., Marcus et al. npj Digital Medicine 2024; Sun et al. Medical Image Analysis 2024; Vorberg et al. MICCAI Stroke Workshop 2025). The authors should include these task‑specific baselines to demonstrate that StrokeTimer truly advances the state of the art. Without such comparisons, the claim of “first fully automatic benchmark” is overstated.  3.Questionable clinical utility of the 4.5–6h window. The intermediate window (4.5–6 h) is clinically ambiguous: guidelines recommend perfusion imaging for patients in this window rather than relying solely on NCCT. The model’s performance on this class is notably lower (AUC 0.54, F1 0.38) and, as shown in Fig. 3, ECMS even degrades it. The paper should discuss whether this window is worth predicting at all, or whether a binary classification (<4.5 h vs. ≥4.5 h) would be more clinically actionable.  4.Lack of external validation. Although the dataset is multi‑center, all experiments use a single internal split (training/validation/test) from the same pooled data. No separate external cohort (e.g., from a different geographic region or acquisition protocol) is used to test generalization. The authors should either perform leave‑one‑center‑out cross‑validation or validate on an independent dataset.  5.Reproducibility concerns. The preprocessing steps (CT‑BET for skull stripping, resampling to 48×256×256) and the two‑stage training procedure (80 + 70 epochs) are described, but hyperparameters (e.g., λ weights in Eq. 2, δ and γ in ECMS) are only partially reported. The code is promised “upon acceptance” – ideally, a public code repository should be provided at submission to enable reproducibility.  6.Writing and presentation. Several minor issues: the abstract mentions “two national cohorts, and ” with missing names; the dataset split uses only 30 validation and 30 test samples, which is extremely small for evaluating a multi‑class model (especially given only 72 samples in the minority class, a test set of 30 may not be statistically stable). The authors should justify the split size or use cross‑validation.

  • 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.

  • 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 addresses an important clinical problem with a well‑designed, multi‑center dataset. The core idea of using disentanglement to handle style variability and contrastive learning for long‑tailed onset windows is sound and leads to measurable improvements over generic baselines. The interpretability results are compelling. However, the methodological novelty is limited, and the lack of comparisons with recent task‑specific models weakens the claim of state‑of‑the‑art performance. The clinical relevance of the three‑window classification is debatable, and the evaluation protocol (small test set, no external validation) needs to be strengthened. Nevertheless, the paper provides a solid engineering contribution and a valuable benchmark for the community. With moderate revisions (adding key comparisons, clarifying the clinical task, and improving validation), the work would be suitable for publication. Therefore, I recommend weak accept, contingent on the authors addressing the above concerns.

  • 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



Author Feedback

N/A




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.

    I agree with the reviews from all the reviewers.



back to top