List of Papers Browse by Subject Areas Author List
Abstract
Accurate lesion segmentation in acute ischemic stroke (AIS) is critical for treatment decision-making, particularly in emergency settings where non-contrast CT (NCCT) is often the only available imaging modality. While most existing studies focus on infarct core segmentation, joint segmentation of infarct core and penumbra under NCCT-only conditions remains largely underexplored. In this work, we propose a collaborative learning framework for joint infarct core and penumbra segmentation on NCCT, leveraging a foundation model as a unified visual encoder and incorporating multi-level clinical constraints to improve robustness and structural plausibility. Specifically, we adopt a DINOv3-ViT backbone pretrained on large-scale natural images and introduce two auxiliary tasks to regularize the segmentation process. First, a clinically-inspired region constraint, {\it{i.e.}}, mismatch ratio, is employed to enforce structural consistency between infarct core and penumbra. Second, a segmentation-guided conditional diffusion module is introduced to learn a perfusion-aware intermediate representation under supervision from perfusion maps, providing stable regional highlighting of perfusion abnormalities rather than voxel-wise reconstruction. Extensive experiments demonstrate that the proposed framework improves segmentation performance and robustness for NCCT-only stroke assessment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1122_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Zzz0251/COMPASS-Net
Link to the Dataset(s)
CPAISD dataset: https://www.kaggle.com/datasets/orvile/cpaisd-acute-ischemic-stroke-dataset
BibTex
@InProceedings{ZhaDon_Perfusion_MICCAI2026,
author = { Zhang, Donghao AND Zhang, Jianhai AND Jiang, Ruizhou AND Jing, Shuang AND Duarte, Kaue Tartarotti Nepomuceno AND Ganesh, Aravind AND Menon, Bijoy AND Qiu, Wu},
title = { { Perfusion Aware Infarct Core and Penumbra Segmentation on NCCT with Region Constraints } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
CoMPASS-Net, a network to segment core and penumbra regions in non-contrast CT scans for ischemic strokes. the paper describes the use of auxiliary task in order to enhance the segmentation accuracy of the model.
- 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 and the exposition is clear and no too dense. The performances shown are clearly better than the proposed baselines, the Idea behind the method is simple but promising. Ablation study is s clear.
- 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 paper do not report code for reproducibility.
- Fig. 1 report some discarded/registered CLS tokens as a step, but never mention in in the paper.
- The DINO conditioning method is not described, from what i understood is by objective minimization only, not a network conditioning.
- Threshold choices look a bit arbitrary, maybe a graph showing a major range of thresholds can help.
- The presence of MLP Mismatch Decoder is not justified enough to me, it rises me a doubt. Why the ratio (Mismatch Loss) can not be derived directly by the output segmentations?
- 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?
Clarity and readability. The topic is interesting and the method is well engineered to produce a segmentation by using less informative data like NCCT. The idea is simple and grounded with the respect of clinical aspects.
- 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
They proposed a collaborative learning framework for joint infarct core and penumbra segmentation on NCCT, using a foundation model (DINOv3-ViT) as a unified visual encoder and incorporating clinical constraints to enhance robustness and structural plausibility.
The method includes two auxiliary tasks: a clinically inspired mismatch ratio to enforce consistency between core and penumbra, and a segmentation-guided conditional diffusion module that learns perfusion-aware representations from perfusion maps.
Experiments show that this approach improves segmentation performance and robustness for NCCT-only stroke assessment.
- 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 organised and easy to follow. Using only NCCT to segment penumbra and core addresses the gap of not having access to perfusion CT or diffusion MR for evaluating stroke. They use the mistmach ratio and Tmax perfusion maps for traning phase.
- 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.
They have used a dataset that includes NCCT and CTP for large vessle occlusion. The authors write that the have coregistered NCCT and CTP data but they don’t mention what registration and how they do it.
The ground truth the use for their segmentation is derived from rCBF and Tmax thresholds which can introduce biases/errors. In the literatur there isn’t consencus on their threshold for defining ischemic core and hypoperfused region.
Why the intensities for NCCT were cut to [0,100] and for Tmax to [0,30], these need to be explained. How these values were chosen?
Have the authors considered traning their network for core and penumbra from the start instead of computing penumbra based on segmented core and hypoperfused region? Wouldn’t that make their framework smaller and also less time consuming?
- 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 provide sufficient information for 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?
I gave a weak accept because the paper addresses an important clinical problem and proposes a well-structured framework with promising results using only NCCT, which is highly relevant in practice. Despite some methodological ambiguities (e.g., registration details, threshold-based ground truth, and preprocessing choices), the overall contribution is solid and shows potential impact. With clearer explanations and minor improvements, the work could be significantly strengthened.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
The paper introduces a multitask framework for ischemic stroke lesion segmentation on NCCT scans, utilizing a DINOv3 foundation model as the backbone for a dual-head UNet to segment the infarct core and total ischemic tissue. To refine these predictions, the model undergoes joint optimization driven by three objectives: a standard segmentation loss, an auxiliary task using an MLP to predict the mismatch ratio (IschemicArea / CoreArea), and a conditional denoising diffusion model tasked with approximating Tmax perfusion maps.
- 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 research addresses a highly relevant, but challenging problem: predicting stroke penumbra and core from NCCT. The methodology is innovative, particularly in its use of a modern foundation model (DINOv3) and the attempt to derive physiological perfusion information (Tmax) from non-functional imaging. Qualitatively, the generated Tmax maps show promise in identifying blood flow abnormalities despite the limited signal in NCCT. The integration of auxiliary tasks to provide a “physiology constraint” is a conceptually sound approach to improving segmentation performance.
- 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 manuscript suffers from significant technical and presentation issues. There are apparent circular dependencies in the proposed pipeline, where the encoder seems to require outputs from modules that have not yet been executed. The terminology is often misleading, specifically the use of “collaborative framework” to describe what is essentially a multitask learning setup. Furthermore, the experimental validation is hindered by a lack of transparency regarding the dataset (missing inclusion criteria, specifications from the scanners, distribution of lesion characteristics) and a weak quantitative analysis. The ablation study shows marginal improvements (e.g., 0.002 in Dice) when adding the high-complexity Diffusion module, which lacks statistical significance testing and calls into question the trade-off between model parameters and performance gains.
- 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 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
General Concerns
- The term “collaborative framework” is misleading. In the context of medical imaging and AI, this often refers to Federated Learning. The proposed approach is more accurately described as a multitask learning framework or a segmentation model with auxiliary optimization objectives.
Introduction
- The second paragraph provides a highly superficial overview of related work. Several relevant state-of-the-art studies that attempt to predict infarct core from NCCT are missing. Authors could also consider mentioning efforts to predict penumbra from other modalities non-functional modalities.
- The authors must expand the literature review to include specific state-of-the-art baselines for NCCT stroke segmentation to better contextualize their proposed method and results. For example, a more detailed comparison with the I2PC-Net model could be helpful.
Methodology and Pipeline
- There are multiple typos in Figure 1: “regin” instead of “region”, “folw” instead of “flow”, and “Supervied” instead of “supervised”.
- Section 2.2 and the Overview are inconsistent. Section 2.2 mentions a dual-head module for probabilistic masks of the core and penumbra, while the Overview states the penumbra and core.
- Regarding the 1.8 mismatch ratio (Section 2.3): In clinical practice, this ratio is calculated using total volumes. Please clarify how this was adapted for 2D axial slices and if this introduces bias.
- In Section 2.4, map “M” is mentioned but not labeled in Figure 1.Additionally, clarify why the output of the Diffusion Decoder is called a “Generative Attention Map” if it is intended to be an estimate of Tmax (noted as Tmax Hat).
Datasets and Implementation
- Section 3.1 lacks essential details regarding the patient cohort. Please provide inclusion/exclusion criteria, study nature (retrospective/prospective), imaging equipment (NCCT/CTP) specifications, image dimensions, lesion characteristics (size, location), and any other relevant information.
- In the Implementation subsection, there is confusion regarding Lambda parameters. The authors mention Lambda 2 increases to prioritize latent representation, but the description suggests they may be referring to Lambda 3 (reconstruction space). Please clarify this.
- The 2D re-implementation of I2PC-Net for a “fair comparison” is questionable. Reducing a 3D method to 2D may strip it of its inherent advantages, potentially biasing the comparison and diminishing the novelty of the proposed method.
- Clarify how the baseline models were adapted to output two maps (infarct and ischemic tissue) for the evaluation.
Results and Discussion
- The description of the ablation study results is vague. Phrases stating that modules “help improve performance” should be replaced with specific quantitative measurements.
- The quantitative gain from adding the Diffusion module is extremely low (e.g., 0.002 in infarct core) and actually results in performance loss in other metrics. Given the massive parameter overhead of a Diffusion model, this requires a much more rigorous justification or a statistical significance analysis to prove the differences are real.
- While the qualitative Tmax maps are interesting, the authors should discuss and justify why the model is being burdened with such a complex module for such marginal quantitative gains.
- 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 core idea of using Foundation Models and Diffusion-based auxiliary tasks for stroke segmentation is novel and addresses a critical clinical need, the current version of the manuscript has significant flaws. The introduction provides a highly superficial presentation of state-of-the-art works, failing to adequately contextualize the proposed method against existing baselines for NCCT stroke segmentation. Furthermore, the dataset description is insufficient for reproducibility, and the quantitative results do not currently justify the high complexity of the proposed architecture. The marginal improvements observed in the ablation study, coupled with the lack of statistical significance testing and missing baseline comparisons in the evaluation, suggest that the paper requires a major revision and more rigorous validation before it is ready for publication.
- 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 authors provided clarification on terminology, corrected several typos, and shared additional details regarding the method and dataset. While they did not provide a response for the state-of-the-art contextualization, the proposed methodology presents a novel and promising approach to a high-impact clinical problem.
Because the technical contribution is significant and addresses a critical area in medical imaging, I have decided to increase my score. However, I strongly recommend that the authors take this opportunity to refine the final version of the manuscript. Properly situating this work within the current landscape of NCCT analysis is essential for the scientific community to fully appreciate its value and novelty. Ensuring a more rigorous contextualization will significantly improve the paper’s presentation and long-term impact.
Author Feedback
We sincerely thank all reviewers for their thorough and insightful comments. Below we address the methodological and presentation concerns raised. Clarification of terminology and implementation: “Collaborative framework” refers to a standard end-to-end multitask framework in this paper. The segmentation branch predicts infarct core and ischemic tissue; the use of “core and penumbra” in Sec. 2.2 was a typo. The map “M” denotes the attention map derived from segmentation outputs for attention gating. In Sec. 2.3, the statement that mismatch is predicted “from segmentation outputs” is inaccurate; as shown in Fig. 1, mismatch is predicted by an MLP from encoder features. In Sec. 3.1, λ2 should be λ3, corresponding to the diffusion loss. These corrections clarify the intended implementation without changing the methodology. For data concern, the dataset is a LVO dataset of 310 retrospective AIS patients, with following detailed inclusion criteria in [1]. Perfusion maps were affine-registered to NCCT space after skull stripping. Intensities were clipped to [0,100] HU for NCCT and [0,30] s for Tmax following [1]. Infarct core and ischemic tissue labels were obtained by standard thresholds (rCBF < 0.3, Tmax > 6s) and clinically reviewed. [1] Cai, Y. et al. CT perfusion map generation from multiphase CTA using a generative adversarial model for acute ischemic stroke. American Journal of Neuroradiology, 46(12), 2535–2544.Fig.1 may have caused confusion. The arrows returning to the DINOv3 encoder indicate gradient propagation during joint optimization rather than recursive execution. The diffusion branch uses segmentation outputs only as conditioning signals, while denoising is performed by an independent diffusion network. The framework follows sequential multitask optimization, and inference requires only NCCT input. For mismatch branch, threshold of 1.8 follows the DEFUSE-3 criterion. The MLP predicts mismatch status from pooled latent features, providing a global structural constraint complementary to pixel-wise supervision, reducing the influence of small lesion fluctuations on ratio estimation and improving robustness in small infarct cases. Although mismatch is clinically volumetric, we adopt a slice-level approximation, where consistently derived 2D annotations serve as region-level supervision. For diffusion branch, it is designed as a physiology-aware auxiliary constraint rather than a Dice-optimization module. Under Tmax supervision, it learns perfusion-related features and generates pseudo-perfusion representations (generative attention map) from NCCT. The modest Dice gain under 1.8 threshold suggests that the segmentation backbone already captures substantial lesion information, while the diffusion branch contributes through representation regularization and physiological interpretability. Although the Dice differences are modest, we note consistent trends across multiple experiments (Table 2, for threshold 0.8 and 2.8, infarct dice gain are 0.016 and 0.009, Ischemic tissue dice gain are 0.010 and 0.008), suggesting the improvements are systematic rather than random. Furthermore, this framework may provide a promising direction in our future work by mismatch-aware regional structural priors within denoising, which may further improve both pseudo-Tmax quality and Dice gain. For fairness, all baselines, 2D or 3D methods, were adapted into the same dual-head 2D setting while preserving their core architectural components, which ensures that performance differences mainly reflect architectural design rather than dimensionality differences. The 2D formulation was adopted due to NCCT characteristics, including anisotropic resolution, ~5 mm slice thickness, limited through-plane continuity, and variable slice counts. Under these conditions, slice-wise 2D modeling is a practical and widely adopted strategy for NCCT lesion analysis. Code will be released upon acceptance.
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.
This paper proposes a multitask framework for infarct core and penumbra segmentation on NCCT using auxiliary physiological constraints and diffusion-based modeling.
Reviewers agree the problem is clinically relevant with a well-structured and potentially innovative approach, but raise concerns about unclear methodological details, weak justification of some design choices, limited transparency in data and preprocessing, and insufficiently supported performance gains. I suggest that the authors address these raised issues in the rebuttal.
- 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.
There is consensus that the paper addresses a clinically important problem and proposes a technically promising multitask framework for infarct core and ischemic tissue segmentation from NCCT. The main concerns have been reasonably clarified in the rebuttal, and I therefore recommend acceptance.
However, there is a valid concern about the limited contextualization against state-of-the-art NCCT stroke analysis methods. I encourage the authors to address this as much as possible in the camera-ready.
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.
The rebuttal does not address point-by-point the detailed observations from Reviewer 3, but the paper nevertheless makes a valuable and clinically-relevant contribution to the field of analysing non-contrast CT scans in patients presenting with stroke to clinics. Several methods are already commercially available for similar tasks, and this work does not seem to greatly improve their performance, but it does opens the avenue of research on the theme using the breadth of CT scans from multiple databases by giving the possibility to researchers worldwide of differentially identifying the core/penumbra infarcted regions, to be able to refine strategies to reduce the area affected by the stroke. In the final version, to carefully address point-by-point the suggestions of Reviewer 3 will be expected.
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 rebuttal has sufficiently addressed the major concerns of the reviewers. The general core methodology is novel and addresses a critical clinical need.