Abstract

The combination of multi-modal medical imaging for ischemic stroke infarct segmentation is crucial for clinical treatment. However, existing methods often improve segmentation accuracy at the cost of efficiency, rendering them impractical for mobile health applications. To overcome this limitation, we integrate Mamba, a state-space model for long-sequence modeling, with convolutional operations to capture both global and local dependencies. To further enhance the feature representation, we incorporate multi-scale feature interaction and frequency-domain processing. As a result, we propose a novel Efficient Frequency-enhanced Multi-Scale Network (EFMS-Net) to achieve an optimal trade-off between segmentation accuracy, inference speed, and parameter efficiency. Extensive experiments on four datasets demonstrate the effectiveness and efficiency of EFMS-Net. We release a new dataset to promote further research in ischemic stroke infarct segmentation. The dataset is available on GitHub.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3331_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YanJie_EFMSNet_MICCAI2025,
        author = { Yang, Jie and Shen, Shaowei and Fan, Xuwei and Chen, Ning and Gao, Zhibin and Huang, Lianfen and Zhan, Yihong},
        title = { { EFMS-Net: Efficient Frequency-Enhanced Multi-Scale Network for Ischemic Stroke Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {164 -- 174}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    the author proposed anew solution to stroke lesion segmentation with a focus on inference speed and efficiency, called EFMS-Net. Particularly, the EFMS-Net combines a Dual-branch Adaptive Frequency Fusion (DAFF) block that merges local conv and global mamba-based features in frequency domain using DCT-based attention mechanism; another contribution is the Laplacian-enhanced multi-scale attention (LMSA) block for capturing spatial-channel dependencies and enhancing edges using LOG operator.

  • 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 experiments on 3 major public datasets + a private dataset are pretty exhaustive in terms of converting different types of MRI for stroke images.
    2. the method shows very good computational efficiency when comparing with standard benchmarks
    3. DAFF is nothing particular but LoG is quite interesting for its effectiveness for boundary details.
  • 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. statistical anaylsis is needed. particularly, the differences across setup in ablation study with very minor performance differences make me wonder if the difference is statistically significant, particularly given the ISLES datasets are relatively small and the performance variance must be not very tiny. This also means I recommend you to report confidence interval alongside the metrics.
    2. discussion of related work and motivation: the authors should discuss frequency-domain segmentation models; the authors should discuss the motivation for stroke segmentation, what’s the clinical necessacity? ISLES seg and ATLAS serve very different purposes. the proposed method does not truly address anything specific to stroke but rather a general modification to existing methods that may rather work on other imaging modalities/organs.
    3. the setting is not very clear. Do you input all the modalities? or just DWI? what b value? or along with FLAIR/ADC. is your ISLES18 input a 2D input with 5 modalities? what did transform did you do to the slice to make them uniform? cropping? resize?
    4. lack of downstream task to increase clinical utility. particularly, for AIS, stroke onset is a critical task using MRI, worth evaluating the performance with segmentation mask.
    5. the visualization analysis is subpar. What’s the model’s failing behavior? how do you prove your claim about better edge segmentation?
    6. The authors only compare some standard approaches as baseline. The manuscript should compare other SOTA works on these 3 public datasets. many of the SOTA works achieve much higher DSC.
  • 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.

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    Please see weakness for rebuttal. I will also reconsider the score depends on the dataset release criteria. “upon reasonable request” is not a proper way to release dataset. direct release with applicable license/agreement is standard of approach.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    it’s a common task that have been done for years and the authors did not really convince me the contribution to this well-established task is significant.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel efficient frequency-enhanced multi-scale network (EFMS-Net), which achieves a better tradeoff between segmentation accuracy, inference speed, and parameter efficiency. Additionally, extensive experiments on four datasets demonstrate the effectiveness and efficiency of EFMS-Net.

  • 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 author proposes the DAFF block to improve segmentation in regions with fuzzy boundaries and low contrast and the LMSA block to enhance edge representation, while also claiming to provide a new dataset for 3D ischemic stroke segmentation.

  • 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 author claims that EFMS-Net can be used for mobile health devices. Please discuss in detail which specific mobile devices EFMS-Net can be applied to and how it functions?

    (2) What data augmentation strategies were used in the article to mitigate the overfitting problem?

    (3) The results in Table 1 show that on the LSLES’18 dataset, EFMS-Net has fewer parameters and FLOPs than LightM, while on the LSLES’22 dataset, EFMS-Net has fewer parameters but more FLOPs than LightM. Why is this the case?

    (4) The results in Table 1 indicate that the parameter counts for UX-Net and LightM are the same across all three datasets, whereas the parameter counts for other methods vary across the datasets. Why is this?

    (5) From Figure 2, the segmentation results of different methods show minimal differences. To facilitate further comparison, please provide corresponding metric scores for each segmentation sample.

    (6) Please conduct comparisons with more lightweight methods and recently published approaches, such as [1], [2], and [3].

    (7) The overall write-up needs improvement.

    References [1] Chen J, Chen R, Wang W, et al. TinyU-Net: Lighter Yet Better U-Net with Cascaded Multi-receptive Fields[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 626-635. [2] Dai D, Dong C, Yan Q, et al. I2u-net: A dual-path u-net with rich information interaction for medical image segmentation[J]. Medical Image Analysis, 2024, 97: 103241. [3] Kuang H, Wang Y, Tan X, et al. LW-CTrans: A lightweight hybrid network of CNN and Transformer for 3D medical image segmentation[J]. Medical Image Analysis, 2025: 103545.

  • 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.

  • 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?

    This paper proposes an efficient frequency-enhanced multi-scale network (EFMS-Net). Additionally, the author claims to provide a new dataset for 3D ischemic stroke segmentation.

  • 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 author has solved my problem.



Review #3

  • Please describe the contribution of the paper

    Authors propose EFMS-Net, a deep network for the segmentation of ischemic stroke lesions. This network leverages the classical UNet structure, where two new blocks are implemented: DAFF (Dual-branch Adaptive Frequency Fusion) and LMSA (Laplacian-enhanced Multi-Scale Attention). The proposed approach was extensively validated with four datasets (three public and one private). The proposed approach was also compared against well-known models in the literature. Results on the four datasets demonstrate superiority of the proposed approach w.r.t. other alternatives in the literature.

  • 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.
    • Extensive validation with public and private ischemic stroke datasets.
    • Extensive comparison w.r.t. state of the art approaches (nnUNet, UNETR, nnFormer, etc.)
    • Results show that the proposed approach achieves the best performance regarding lesion segmentation, with models with less parameters and FLOPs.
    • Ab ablation study allows estimating the contribution of each one of the methodological components.
  • 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.
    • Major clarity is needed regarding the training and validation of models: Authors mention that they used a 5-fold cross-validation. In total, including training and testing cases, ISLES22 has 400 and ISLES18 has 156, which is the same amount reported in the experimental setup. Nevertheless, to my knowledge, the test cases are hidden (cannot be accessed), did authors forgot to tell that they only used the training cases?
  • 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.

    (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?

    In general, the paper presents a novel and well-validated approach, tested on both public and private ischemic stroke datasets, with extensive comparisons against strong baselines. It achieves state-of-the-art performance in lesion segmentation while using fewer parameters and FLOPs. The only clarifications needed are regarding the use of test data in cross-validation, but this does not significantly undermine the strong experimental setup and results.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The authors’ feedback satisfactorily addressed my sole concern regarding the data used for training. Consequently, my decision remains to accept this work.




Author Feedback

We sincerely thank the reviewers for their valuable feedback and insightful suggestions. We address their main concerns below and will add the GitHub link in the final version for the direct release of the self-made dataset under the applicable license/agreement.

  1. Confidence intervals and the setting (R1Q1Q3&R3Q2): the tailed t-Test results (P-value) were calculated across modules but omitted due to space limitations, indicating statistically significant differences. We feed 3D volumes of all modalities directly into the model as multi-channel inputs, e.g., using 5 channels (CBF, CBV, MTT, Tmax, CT) for ISLES18. During each epoch, the framework automatically crops fixed-size patches for each case. To enhance model performance, data augmentation techniques such as random Gaussian smoothing, scaling, flipping, rotation, and additive Gaussian noise are applied.
  2. Related work, motivation and clinical necessity (R1Q2&R3Q1): Frequency-domain segmentation models(Refs. 12–14)convert spatial data into the frequency domain to effectively extract discriminative features. E.g., FcaNet extends channel attention to the frequency domain, leveraging multiple frequency components to enhance feature embedding and representation. AI-assisted stroke segmentation is vital for accurately delineating infarct areas, guiding timely interventions, monitoring treatment efficacy, and predicting recovery. Multi-modality fusion can further enhance segmentation performance by providing complementary information. These points are also mentioned in the Introduction and will be refined later. Four datasets include diverse infarct core characteristics, including different stages, sizes, locations, and imaging modalities. Therefore, the research based on these datasets is valuable. To address challenges of blurred boundaries and irregular shapes in stroke lesions, the model incorporates Frequency Attention and Laplacian of Gaussian modules to enhance edge information from both frequency and spatial perspectives. Its lightweight structure allows for deployment on mobile medical vehicles or edge computing devices, enabling rapid segmentation by remotely receiving CT/MRI data or direct access to hospital imaging systems, thereby aiding in urgent medical decision-making. Furthermore, the proposed model demonstrates promising performance across other imaging modalities and organs.
  3. Result analysis (R1Q4-6&R3Q3-6): for example, in ATLAS v2 dataset shown in Fig. 2, both cerebrospinal fluid and lesions exhibit low signal intensity in the T1-weighted modality, making it more difficult to distinguish lesions. EFMS-Net performs better in boundary evaluation and regional similarity, while other methods suffer from relatively severe over-segmentation. According to Tab. 1&3, the baseline of EFMS-Net uses standard and transposed convolutions for downsampling and upsampling, while LightM employs MaxPool, interpolation, and depthwise separable convolutions. Consequently, when processing the larger input volume of the ISLES22 dataset, the FLOPs increase more significantly for EFMS-Net. Despite incorporating DAFF and LMSA to mitigate FLOPs, the increase remains noticeable. The parameters of both UX-Net and LightM vary slightly across three datasets and we report only the values rounded to one decimal place. Due to conference regulations, we cannot disclose the specific values now. We will add metric scores for each sample in Fig. 2. The selected competing methods are representative, strong baselines in segmentation tasks(R2). To ensure a fair comparison, we conducted all experiments under the same configuration. Experiments further confirmed that our method consistently outperforms LW-CTrans, I2U-Net and TinyU-Net across all evaluation metrics. Unlike I2U-Net and TinyU-Net, which employ 2D image processing and thereby lose spatial information, our model directly processes 3D images.
  4. Public dataset (R2): yes!we only used labeled training cases for 5-fold cross-validation.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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”.

    N/A

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    This work proposes a new network architecture for brain lesion segmentation. The method is well motivated, and the experiments showed the effectiveness of the method well. One reviewer has some concerns on the compared methods and discussion on motivation, but overall the works is above the acceptance threshold.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    This paper has received mixed scores, which are maintained after the rebuttal process. I find the submission to be really on the fence, with no major reasons to either reject it or accept it. While the task is widely studied, and the added components exist in the literature (being merely a plug-in of different components), the empirical validation is somehow extensive (at least given the length constraints), with three different datasets and multiple baselines, showcasing the superiority of the proposed approach consistently. Based on these comments, I am weakly inclined towards acceptance.



back to top