Abstract

Acute ischemic stroke is one of the major causes of mortality and disability worldwide. Although thrombectomy is an effective intervention, it carries a lot of risks such as hemorrhage and vascular injury. Thus, it is crucial to accurately predicting postoperative infarct before intervention, providing the guidance for treatment. The existing perfusion imaging techniques relying on fixed thresholding approaches mostly fail to account for individual differences in collateral circulation recruitment, which has been proven to effectively reflect infarct severity. In this work, we take the first step toward integrating collateral circulation status into deep neural network, enabling the model to learn and capture hemodynamic cues for infarct prediction. Specifically, we establish the first brain computed tomography perfusion (CTP) dataset including collateral circulation status and further conduct a thorough analysis of its effectiveness in predicting infarcts. Based on the findings, we propose a novel multi-modal fusion module that integrates spatiotemporal features of multiple modalities. Specifically, a bi-directional Mamba structure is developed to extract the sequential information, which is then fused with collateral priors via a mixture-of-experts mechanism. In addition, a two-stage infarct prediction module is developed to successively localize and segment the infarct region under the guidance of collateral circulation status. Finally, both infarct localization and segmentation performance of our method are validated to outperform 14 state-of-the-art methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0958_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Frankenstein2026/CCGM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GuoYic_Collateral_MICCAI2025,
        author = { Guo, Yichen and Zhao, Xinyi and Dai, Lisong and Dong, Heming and Jiang, Lai and Xu, Mai and Li, Shengxi},
        title = { { Collateral Circulation guided Multi-modality Fusion Network for Postoperative Infarct Prediction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15974},
        month = {September},
        page = {95 -- 105}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes for the first time a collateral circulation-guided multimodal fusion network (CCGM). By constructing the first CTP data set (CTPPC) containing collateral circulation status and designing a bidirectional Mamba structure and an expert mixing mechanism, it realizes accurate localization and segmentation of postoperative infarct areas and significantly surpasses the prediction performance of 14 existing methods.

  • 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 first quantitative collateral circulation dataset (CTPPC): To construct the first CTP dataset integrating multi-phase collateral circulation diagrams (ART/VRT/LRT), dynamic collateral diagrams were generated by contrast agent distribution differences, and it was verified that it was strongly correlated with infarct area (r = 0.82), breaking through the limitations of individual differences of traditional fixed threshold methods.
    2. Bidirectional Mamba spatiotemporal modeling: Innovative use of Mamba structure to extract the long-term spatiotemporal dependence of CTP multi-phase scanning, which improves feature coherence compared with traditional RNN/CNN, and the inference speed is 1.8 times faster at 128 × 128 resolution.
    3. Parameter graph-driven expert hybrid fusion (MoE): Using perfusion parameter graph as a domain knowledge expert, a dynamic gating weight allocation mechanism is designed to improve the CBF/CBV/MTT/Tmax modal fusion Dice by 4.2% and solve the problem of multi-modal information conflict.
  • 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 manuscript lacks corresponding visual comparative experiments to clearly demonstrate the effectiveness of the proposed methods. 2.There is an absence of experimental comparisons with recent state-of-the-art approaches, which limits the assessment of the method’s relative performance. 3.It remains unclear whether the experimental data are derived from multi-center sources, which is essential for evaluating the generalizability of the method. 4.Key implementation details, such as hardware configuration and batch size, are not specified, making it difficult to assess the reproducibility and computational feasibility of the experiments.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Although the integration of collateral circulation into the deep learning framework is innovative, the study suffers from insufficient experimental validation and a lack of detailed methodological descriptions. Addressing these issues would significantly enhance its impact in translational stroke research.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors aim to localize and segment the infarct region from brain CTP data. A brain CTP dataset was acquired which consists of 15 sequential scans in time. Derived from the CTP scans, the authors also compute 4 parameter maps and 3 collateral maps. This dataset is claimed to be published. The authors propose a multi-modality fusion module, which employs a Mamba-based feature extraction from the CTP scans and an experts-oriented feature fusion with the parameter maps. The obtained multi-modal features and the collateral maps are then provided to a two-stage conditional network, where the infarct region is localized before the cropped image region is provided to the segmentation module to obtain the final result.

  • 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 authors provide a strong evaluation. For segmentation, they compare to 11 different works on six metrics. For localization, they provide a comparison to three methods on one metric.
    2. Further, the ablation study assesses the performance when removing modalities as well as when removing individual components from their network.
    3. Overall, their method shows great performance when compared to related work and their ablation study justifies each component of the arguably complicated model design.
    4. The authors acquired an interesting brain CTP dataset which they claim to publish.
    5. The paper is well written and mostly easy to follow. The methodology is rather complex, however, Fig. 2 provides a good overview.
  • 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. I have one major concern regarding the evaluation. Namely, it is unclear if related work methods only received the CTP scans as input, or if they (same as the proposed method) received CTP scans, parameter maps and collateral maps, e.g. by concatenation in the channel dimension. This needs to be clarified in the manuscript.
    2. For this point I am assuming that related work received only CTP scans as input. In this case, the comparison in Table 1 is not fair. Figure 3 a) shows that the proposed method when only using CTP scans achieves a Dice score of 53.19. With this the proposed method still outperforms related work, however, much closer compared to what is reported in Table 1 and scores for other metrics are not available in the manuscript. Thus, for a fair comparison in this case, I would strongly suggest the authors to provide CCGM results using only CTP data for all metrics in Table 1and emphasizing clearly, that CCGM uses additional parameter and collateral maps in contrast to related work.
  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors contribute a dataset and propose an interesting method with a strong evaluation. However, the authors did not specify whether related work methods received only CTP scans or also the parameter and collateral maps in their evaluation. This might impede the performance improvements of their method compared to related work and needs to be addressed in the manuscript.

  • 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

    The paper presents a novel dataset based on CT perfusion scans. The dataset includes pre-operative CT perfusion sequence scans, perfusion parameter maps, and collateral status maps as the difference between CTP scans. The dataset includes manually delineated infarct segmentations from post-operative data. The authors present a novel network architecture for input data and feature fusion for prediction of post-operative infarct segmentation and localization. The results of this method show considerable improvement compared with a large number of medical image segmentation networks, and three localization networks.

  • 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.
    • Imaging techniques, data acquisition, and dataset generation are concisely and clearly explained.
    • Descriptive graphics and figures.
    • Extensive and clear description of the network and their components.
    • Designed architecture is sophisticated, carefully inputting and fusing the multiple types of images. The effect of each network component is tested in an ablation study.
    • Large comparison study against medical image segmentation and localization networks, proving segmentation and localization improvement.
    • Strong story line and results: creation of the dataset including very relevant clinical information, great network design based on preliminary data analysis.
  • 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.
    • Collateral and parameter maps should have been briefly described in the introduction since it is not obvious what they are, and it is relevant for the paper.
    • The data analysis results take too much text space in the paper and should have been summarized.
    • Unclear and insufficient details provided in data split for training, testing, and validation necessary for reproducibility. More details should have been provided on how the training / test / validation split was done. Is it done at the patient level? How many patients in each set?
    • Insufficient details provided in comparison against other networks necessary for reproducibility. Are the comparator networks trained using the same input data used for training as CCGM? If so, how is the data input? Are the different input images concatenated in the channel dimension?
    • Since this method can potentially be used for aiding clinical decision making for the treatment of stroke, which a very time-sensitive process, I believe that inference time should have been provided. Including the time to process the CTP to acquire parameter and collateral maps.
  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    • It would have been interested to have a brief description of how pre-operative prediction of infarct size might affect clinical decision-making for the treatment of the stroke. Are there preferred therapeutical approaches depending on the predicted size of the infarct? How do doctors currently evaluate what approach to take without this information? How would this improve the patient outcome?
    • It would have been interesting to see what the increased computational cost (in terms of training and inference time) is compared to the other networks.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a novel dataset and novel architecture that leverages the dataset design for improved post-operative infarct prediction. The paper is written in a clear and concise way. However, it lacks certain details (e.g. data set split and data used for comparator networks) that need to be clarified and added to the manuscript before publication.

  • 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

  1. Visual comparison. [R1] Thanks for the suggestion. Some examples of visual segmented results from our and all the compared methods will be added in revision.

  2. Fairness of comparisons. [R2/R3] Actually, for fair comparison, the compared methods (Attention UNet, Transunet, etc.) combine the parameter maps and the collateral maps with CTP scans as the inputs, via concatnenation in channel dimension. Therefore, the comparisons are fair in terms of the concerned input data, which will be clarified in revision. Besides, as required by R2, we also report the results of our CCGM with the input of only CTP scans on more metrics here, as the supplement results of Fig.3-(a). As a result, the accuracy, precision and recall degrade by 0.94%, 3.12% and 2.2%, respectively. Besides, the HD95, ASSD increase by 1.80mm and 1.13mm, respectively. This just further validates the effectiveness of parameter maps and collateral maps.

  3. More details. [R1/R3] 1) Clinical context. [R1] The data is collected from two top hospitals in China, which will be made public in camera ready version. Therefore, the data is derived from multi-center sources and the generalizability of the method can be ensured. 2) Implementation. [R1] The experiments are conducted with AMD EPYC 7542 32-core CPU and an NVIDIA RTX 4090 GPU, with a training batch size of 8. More reproduction details will be added in revision. 3) Data. [R3] As claimed in Sec 4.1, the training and test data are randomly split as a ratio of 4/1. Actually, the collected 2D slices from the hospital lacks the patient information. Therefore, the data cannot be split at patient level. Note that although in this case, our method outperforms all the compared methods under the same experimental settings. Besides, following 5-fold cross-validation, the random split of train/test has little to no impact on the results. In future work, we will try to ask the hospitals to complete the patient information.

  4. Inference time. [R3] As suggested, we test our inference time. Note that the generation of parameter maps and collateral maps are simultaneously operated, which takes about 1~2s and 0.5~0.85s per slice, respectively. Besides, both the physicians and our method need to leverage the parameter maps for diagnosis. Therefore, the extra inference time of our method only includes the model inference time, which takes 0.14s per slice and is fast enough for aiding clinical decision making for the treatment of stroke.

  5. Recent SOTA methods. [R1] Actually, among all the compared methods, 4 of them are published after 2023. For instance, SiNGR and Bi-JROS are both published in 2024. Besides, comparison with more recent methods such as CCViM (TMI 2025) and WMHCA (Information Fusion 2025) will be added with the released codes on Github.

  6. Description of parameter and collateral maps. [R3] Thanks for the suggestion. Actually, the parameter maps reflect the detailed information on brain tissue perfusion, including cerebral blood flow, cerebral blood volume, mean transit time, and Tmax. Besides, collateral maps refer to a visual representation of brain tissue perfusion status, identifying areas with insufficient collateral supply. The corresponding description will be added in the 3rd paragraph of introduction.

  7. Simplification of analysis. [R3] As suggested, the data analysis results will be simplified in revision. Specifically, the size of Fig.1 will be scaled down. Besides, the text description of Fig.1-(b)/(c) will also be simplified. For instance, in Finding 2, the result description can be simplified into “As shown in Fig.1-(c), together with the CTP scans, the collateral maps tend to bring significant performance gain, which is higher than that of the parameter maps.”




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The manuscript describes a collateral‐circulation–guided fusion network built on the new CTPPC dataset for accurate postoperative infarct localization and segmentation.

    The paper has received 2 weak accept and 1 accept, so there is consensus among the reviewers.

    While reviewers appreciated the innovative idea of integration of collateral circulation into infarct predict, and the construction of a novel dataset, they also mentioned several concerns, which should be addressed in the next revision if possible. Main suggestions include: 1) lack of visual comparison, 2). fairness of method comparisons, 3). lack of clinical context, data, and methodologica/implementation details, and 4). inference time.

    Those comments are recommended to be addressed, most of which should be relatively straightforward.



back to top