List of Papers Browse by Subject Areas Author List
Abstract
Accurate delineation of ischemic core and salvageable penumbra is critical for guiding reperfusion therapy in acute ischemic stroke (AIS). While CT perfusion (CTP) provides quantitative hemodynamic parameters for tissue-window assessment, its availability remains limited in many emergency settings. In contrast, non-contrast CT (NCCT) is nearly universally accessible but lacks explicit perfusion information, making core and penumbra segmentation highly challenging. We study stroke tissue-window segmentation under missing perfusion, where only NCCT is available at inference. In this setting, specialist NCCT models struggle due to subtle contrast and limited supervision, while reconstructing CTP from NCCT constitutes an ill-posed inverse problem prone to hemodynamic hallucination. To address this, we propose \textbf{ProPhyT}, a prototype-based physiological transfer framework that injects CTP-derived hemodynamic semantics into NCCT-based segmentation without reconstructing perfusion maps. ProPhyT consists of three processes: (i) Hemodynamic Prototype Support Bank Construction, where a self-supervised CTP encoder learns a hemodynamic embedding space and CTP features are clustered to form representative physiological prototypes; (ii) Cross-Modal Prototype Retrieval Learning, where an NCCT encoder is aligned to the CTP embedding space to enable prototype retrieval from NCCT alone; and (iii) Physiological Prototype-Conditioned Segmentation, where retrieved prototypes are converted into similarity-aware dense prompts to parameter-efficiently fine-tune a medical foundation model for core and penumbra segmentation. Extensive experiments demonstrate consistent improvements over top-performing methods, supporting more reliable reperfusion decision-making in CTP-limited clinical settings. Code will be available upon acceptance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1889_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/CUHK-BMEAI/ProPhyT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DanYin_Prototypebased_MICCAI2026,
author = { Dan, Ying AND Xu, Zhe AND Zhou, Longxi AND Zhang, Lu AND Ma, Xiangyuan AND Yuan, Yading AND Tong, Raymond Kai-yu},
title = { { Prototype-based Physiological Transfer Enables NCCT-only Hyperacute Stroke Tissue-Window Segmentation Under Missing Perfusion } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed a prototype-based physiological transfer framework that injects CTP-derived hemodynamic semantics into NCCT-based segmentation without reconstructing perfusion maps. They do this in three steps: (1) using a self-supervised CTP encoder to learn hemodynamic embedding space and CTP features that are clustered to form representative physiological prototypes; (2) an NCCT encoder is aligned to the CTP embedding space to enable effective prototype retrieval from NCCT alone; and (3) retrieved prototypes are converted into similarity-aware dense prompts to parameter-efficiently fine tune a medical foundation model for core and penumbra segmentation.
- 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 way they used hemodynamic info of CTP to make a bank and be able to use only NCCT for penumbra and core segmentation in AIS and this is a novel goal. generated from both CTP and NCCT.
- 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 different datasets which in general is good. They have used two modalities (CTP and NCCT) each for a specific step in their approach and then integrated the output of each step into their next step (while using another dataset with different acquisition protocol). For example, in a they make hemodynamic prototype bank from CTP coming from UniToBrain and ISLES2024.Then they use ISLES (CTP + NCCT) for cross-modal retrieval learning. They haven’t clearly explained what may cause problem here. They were not careful in pointing out the temporal resolution of each of these datasets.
Have they considered that given stroke is a dynamic process and changes with time, temporal resolution of CTP matters as well as speed and type of the contrast dye. Also, they need to explain what the purpose of CTP encoder is to reconstruct CTP from masked CTP. This makes their paper unclear.
UniToBrain dataset: CTP acquisition parameters were as follows: Scanner GE, 64 slices, 80 kV, 150 mAs, 44.5 sec duration, 89 volumes (40 mm axial coverage), injection of 40 ml of Iodine contrast agent (300 mg/ml) at 4 ml/s speed. In ISLES 2024, CTP acquisitions are not standardized but typically follow clinical stroke imaging practices. Scanners include Siemens (Somatom Force, AS+, etc. ) and Philips (Brilliance 64, Ingenuity), about 40–60 sec duration, with roughly 30–90 temporal volumes, and cover a limited brain slab of around 40–80 mm depending on the scanner. Acquisition settings vary across centers, but commonly fall within 70–100 kV tube voltage and 100–200 mAs, with contrast injections of approximately 40–60 ml iodine at 4–6 ml/s.
Lack of literature review on related work: for example, articles that have used 4D convolutional neural networks on 4D CTP like “CT Perfusion is All we Need: 4D CNN Segmentation of Penumbra and Core in Patients With Suspected Acute Ischemic Stroke” by Tomasetti et al.
- 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 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
In figure 1, where those scan times for NCCT and CTP come from?! The values have wrong unit; they should be in seconds not in minutes.
CPAISD data includes 112 cases and authors write “official split, 92/10/10 cases are used for training/validation/testing. ” There is an error here.
ISLES 2024 only has final infarct region that authors used as a proxy for the core, which can be wrong since we know stroke is a dynmic process and core at the initial stage is different with what ends up as final infarct obtained from DWI taken after treatment. This lesion is not a good proxy for core.
Tau is not defined in the experiment section!
- 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?
My recommendation is mainly based on concerns about the paper’s methodological clarity, clinical validity, and experimental rigor, despite the fact that the topic is interesting and the use of multiple datasets and modalities is potentially valuable.
A major factor is that the paper combines datasets with different acquisition protocols without sufficiently discussing the implications. This is especially important for CTP, where temporal resolution, scan duration, scanner type, and contrast injection protocol directly affect the hemodynamic information being modeled. Since stroke is a dynamic process, differences in temporal sampling and contrast kinetics across datasets can substantially influence the learned representations and limit the validity and generalizability of the proposed framework. The paper does not adequately acknowledge or analyze these issues.
Another important concern is the lack of conceptual clarity in the proposed pipeline. In particular, the role of the CTP encoder and the purpose of reconstructing CTP/NCCT from masked CTP/NCCT are not clearly justified. The absence of a clear explanation makes the overall approach difficult to interpret.
I was also concerned about that the paper appears to use final infarct regions from ISLES 2024 as a proxy for ischemic core, but this is problematic because final infarct after treatment is not equivalent to baseline core. Stroke lesions evolve over time and treatment can alter tissue outcome, so this choice introduces uncertainty into the labels and may undermine the clinical meaning of the results.
In addition, the paper has weaknesses in scholarship and presentation. The related work discussion is limited and does not adequately position the paper relative to prior studies that directly use 4D CTP for stroke lesion segmentation. There are also several presentation issues and inconsistencies, including incorrect time units in Figure 1, an apparent error in the reported dataset split, and missing definition of the parameter τ in the experiments section. These issues reduce the overall clarity and suggest insufficient attention to details.
- 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 #2
- Please describe the contribution of the paper
This work proposes core and penumbra stroke lesion estimation from NCCT scans via a novel hemodynamic knowledge transfer from CTP. Unlike prior work, which aimed to directly estimate (ctp-derived) perfusion maps from NCCT, authors use a ‘bank’ of physiological patterns to support the NCCT-based lesion estimation.
- 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 full proposed framework is interesting, meticulously designed, and novel. I compliment the authors on the idea behind it.
- Large performance gains compared to prior baselines are reported.
- 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.Experimental Design and Statistical Significance. While the proposed framework introduces a novel concept, the current experimental design presents a significant limitation regarding the data utilization. Although the study leverages a substantial cohort of 367 scans from three datasets (including multi-center data), the final results are reported on a subset of only 10 scans (approx 2.7% of the total data). This restricted evaluation significantly limits the generalizability and trustworthiness of the findings. To ensure a robust assessment of the model’s performance, please provide cross-validated results, at minimum, over the CPAISD dataset.
2.Baseline Comparison and Training Methodology. The manuscript lacks a detailed explanation regarding the training protocol for the comparative baselines (e.g., nnU-Net). Were they trained on the AISD datasets only (92+10 scans)? It is essential to clarify these training conditions and discuss how potential imbalances in training data size (proposed method using 357 scans) may have influenced the reported comparative results.
3.Data Heterogeneity and Lack of Harmonization. The authors utilize perfusion maps from UnitoBRAIN and ISLES’24 without implementing any harmonization protocols, despite these maps being generated by different software platforms. Vendor-specific variability is a well-documented factor in the field that can meaningfully alter perfusion indices (see Charatpangoon et al., 2026). Please address how this lack of harmonization impacts the cross-modal prototype retrieval learning and the structural integrity of the ‘bank’ construction.
Charatpangoon, Pattarawut, et al. “Development and Validation of a Harmonization Pipeline for Multi-Site Computed Tomography Perfusion Cerebral Blood Flow Maps—An Analysis of the AcT Trial.” Clinical Neuroradiology (2026): 1-14.
- 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.
- 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?
The novelty of the work is noteworthy; however, the current experimental design presents significant limitations. Given the current evaluation strategy, the results cannot yet be considered reliable or properly contextualized within the state-of-the-art.
- 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
The authors propose a novel methodology for estimating the infarct core and ischemic penumbra segmentation maps directly from Non contrast CT (NCCT) images. To overcome the inherent lack of blood flow dynamics information in NCCT scans, the method aligns the latent representation spaces of NCCT and CT Perfusion (CTP) modalities. This alignment is enforced through structural, distribution, and instance-level losses, and utilizes k-means clustering to extract representative prototypes from the embedded space. Finally, a physiological prompt is constructed based on the similarity between the NCCT features and the CTP prototypes. The top-k most similar pairs are used to generate a representation that is injected into the intermediate layers of a MedSAM architecture, alongside the structural NCCT information, to perform the final segmentation.
- 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 addresses a highly relevant clinical challenge with a robust, well-justified computational pipeline. The paper is well-written, offering a clear taxonomy of existing strategies and successfully identifying literature gaps that logically lead to the proposed solution. Methodologically, the decision to integrate information from multiple embedded prototypes based on top-k cosine similarity is a highly effective way to handle potentially noisy patient-specific prototypes. Furthermore, the experimental section demonstrates significant performance improvements over both stroke-specific methods and general state-of-the-art segmentation architectures, supported by a comprehensive ablation study that clearly validates the utility of each component.
- 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 primary weaknesses lie in the presentation of certain architectural details and minor reporting omissions in the experimental section. Specifically, there is ambiguity between Figure 2 and the text in Section 2.1 regarding whether the pre-training pipeline utilizes an encoder-only architecture or an encoder-decoder architecture (like SparK). Additionally, in the comparison study, there is a missing reference ([8]) in the results table that is actively discussed in the text, and a lack of detailed explanation regarding the implementation and subsequent poor performance of specific baseline prompting methods (“FT-Auto”).
- 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
- Methodology: There is a lack of clarity regarding Figure 2 and the pre-training pipeline. The diagram suggests that only decoders are used for the self-supervised pre-training process; nevertheless, SparK utilizes both an encoder and a decoder. Please clarify the exact architectural components used in the pipeline to resolve this confusion.
- Experiments: In the “Comparison Study” subsection, the text references “Perfusion Synthesis-based approaches” (citation [8]) to point out their limited performance and difficulties. However, this reference/method is entirely missing from the corresponding results table. Please update the table to reflect the text.
- Experiments: The comparison study is generally well-written, but the prompt methods utilized require further explanation. In particular, it is not clear how “FT Auto” was implemented, nor is it explained why this specific approach degraded the architecture’s performance so significantly.
- 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 authors present a highly innovative and clinically valuable contribution to the field of ischemic stroke lesion segmentation. Attempting to bridge the modality gap between NCCT and CTP by aligning their latent spaces and leveraging multi-modal representation learning is a sophisticated approach. Moreover, the proposed approach is novel, the experimental results are convincing, and the ablation study thoroughly justifies the proposed pipeline. The weaknesses identified are strictly related to presentation, architectural diagram clarity, and minor reporting omissions, all of which are easily correctable.
- 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 have addressed my primary concerns regarding architectural ambiguity and experimental reporting. Specifically, the clarifications on the SparK-based pre-training pipeline and the inclusion of missing references in the comparison tables significantly improve the manuscript’s transparency.
This work tackles a highly relevant clinical challenge with a well-explained and solid methodology. The proposed approach is particularly noteworthy as it offers a promising alternative to the traditional paradigm of synthetic modality generation followed by segmentation. Additionally, given the clarity of the presentation and the importance of the problem, I believe this work will be a valuable contribution to the conference.
Author Feedback
We are glad that reviewers find our work to address a “highly relevant clinical challenge” (R3) with an “interesting, meticulously designed, and novel” method (All), “significant improvements” (R2/R3), “convincing results and comprehensive ablations” (R3). We prioritize the major concerns below.
Q1 (R2): Evaluation size and baseline fairness. A1: We used the same official CPAISD split (92/10/10 cases) for all baselines and CPAISD is the only target benchmark for NCCT-only hyperacute tissue-window segmentation with CTP-derived core/penumbra labels. Thus the test set is not 10/367: UniToBrain and ISLES are just auxiliary sources for prototype construction/alignment without extra target labels, while target evaluation follows the official 10/112 CPAISD test set. We also note R2 likely meant CPAISD, not AISD. Yet, we agree that stronger validation is useful and have added 5-fold cross-validation on CPAISD: FT-P-ProPhyT obtains 71.03% Avg. Dice vs 21.20/23.60/48.52/67.50 for nnUNet/nnUNet(Syn.)/BEA-Net/SAM-Med2D(FT-P).
Q2 (R1/R2): CTP heterogeneity/harmonization and temporal protocol differences. A2: We agree that CTP acquisition/software variability is important. While we do not explicitly model such heterogeneity, ProPhyT does not directly operate on raw 4D CTP temporal curves. Instead, it learns high-level embeddings from processed 3D CTP perfusion maps, where temporal dynamics have been summarized into physiological parameters such as CBF and Tmax, and retrieves similarity-weighted prototype patterns for robustness to cross-site variations. Top-K aggregation further reduces reliance on any single noisy prototype, consistent with our ablations and R3’s positive assessment. Importantly, we evaluate on the independent CPAISD (NCCT-only, not used in prototype construction/alignment), where strong performance suggests good practical generalization of the learned physiological embedding and retrieval pipeline.
Q3 (R1/R3): SparK masked reconstruction and role of the CTP encoder. A3: SparK is used as a masked image modeling-based self-supervised learning method, where representations are learned by reconstructing hidden regions from visual context without external labels. During pretraining, SparK includes an encoder and a lightweight decoder; the decoder supplies the reconstruction loss and is discarded afterward. Only the pretrained CTP encoder is retained to define the hemodynamic embedding space and construct the prototype bank. The pretrained NCCT encoder is then aligned to this space for prototype retrieval. We have revised Fig. 2 to show the decoder as a dashed pretraining-only branch.
Q4 (R1): ISLES final infarct as proxy core. A4: We agree that follow-up final infarct is not equivalent to baseline ischemic core. Thus, ISLES labels are not used as target ground truth; CPAISD is the only dataset for supervised core/penumbra training and evaluation. In ISLES, the final-infarct mask is used only during NCCT-CTP retrieval learning as a weak outcome-derived lesion prior: representations of lesion-similar cases are pulled closer, while unrelated cases are separated. This makes cross-modal alignment lesion-aware and reduces domination by normal/background similarity, without treating final infarct as an exact core label. We will revise the wording.
Q5 (R1/R3): Clarifications on CTP literature, timing, split, τ, baselines, and FT-Auto. A5: We will add Tomasetti et al. and clarify that 4D CTP methods require CTP at inference, while ProPhyT targets NCCT-only missing-perfusion inference. We will correct Fig. 1(c) to overall examination time, not gantry scan duration. The split is correct: 92+10+10=112.We set τ = 0.07.We rename Syn. rows as Pix2Pix+nnU-Net/Swin-UNet with [8]. We have explicitly clarified on Page 7 that detector-derived coarse prompts correspond to our FT-Auto variant, avoiding confusion with the interactive mode.
Other minor concerns will be carefully revised. Code will be released to help follow the details.
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 prototype-based physiological transfer framework to enable infarct core and penumbra segmentation from NCCT by leveraging CTP-derived representations.
Reviewers agree the problem is clinically important and the idea is novel with promising results, but raise concerns about experimental design (limited evaluation subset and unclear baseline setup), handling of heterogeneous multi-dataset data, and conceptual and methodological clarity (e.g., role of CTP encoder and ground truth definition). I suggest that the authors address these points 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.
The reviewers generally recognize the novelty and clinical relevance of this work. In my view, the rebuttal sufficiently clarifies the main concerns. I therefore recommend acceptance.
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 has sufficiently addressed the major concerns of baseline setup, methodology clarification, data heterogeneity. Overall, the method is interesting for an important clinical problem.
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.
agree with the majority of reviewers