Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

The domain gap caused by variations in image appearance is a primary obstacle to transferring medical image segmentation models from laboratory research to clinical practice. To address this issue, domain generalization techniques have been introduced to enhance the transferability of the model. However, due to the static nature of existing perception and feature extraction mechanisms, current methods still lack sufficient adaptability in practical generalization scenarios. Inspired by the highly generalizable dynamic information focusing and fusion capabilities of the human visual system, this paper proposes the Self-Coordinated Focusing Network (SCFNet), a novel model designed for modality-agnostic medical image segmentation with domain generalization ability. The model comprises two core components: Dynamic Focusing Convolution (DFC) and the Interactive Attention Fusion Decoding (IAFD) module. DFC dynamically adjusts convolutional aggregation weights and receptive-field scales by pre-perceiving target types and their spatial scale distributions, enabling input-adaptive, target-aware feature perception. Furthermore, IAFD enhances feature fusion efficiency and information interaction during decoding by decomposing semantic information and introducing interactive attention across semantic hierarchies. We evaluate SCFNet on ten datasets spanning four imaging modalities. Experimental results demonstrate that SCFNet consistently outperforms multiple state-of-the-art methods in multi-modal settings, exhibiting superior segmentation accuracy and stronger generalization capability.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1055_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{TanWen_ModalityAgnostic_MICCAI2026,
        author = { Tang, Wentao AND Deng, Hongmin AND Qing, Linbo AND Tan, Rui},
        title = { { Modality-Agnostic Domain Generalizable Medical Image Segmentation via Self-Coordinated Focusing } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a modality-agnostic network for medical image segmentation with the ability to generalize across different domains. The model includes a Dynamic Focusing Convolution (DFC) component and an Interactive Attention Fusion Decoding (IAFD) component to achieve this goal.

  • 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.
    • Applicable to various modalities including dermoscopy, radiology, ultrasound, and colonoscopy within a framework.
    • Comparing with state-of-the-art segmentation methods across multi-modal datasets.
  • 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 sensitivity to hyperparameters, which is an important factor in maintaining stable performance in cross-modality and cross-domain segmentation, is not investigated.
    • To truly generalize, the model needs diverse training data from multiple modalities and domains. Otherwise it may overfit to some modalities.
  • 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
    • Figure 2 contains too many elements, making it difficult to follow and understand.
    • The authors are encouraged to investigate sensitivity to some hyperparameters.
  • 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?

    Despite a few limitations, the paper presents an interesting approach and offers useful insights into a segmentation method that can be extended to multiple modalities. It also demonstrates good accuracy.

  • 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.

    After reading the rebuttal, I found the authors’ response convincing.



Review #2

  • Please describe the contribution of the paper

    This paper proposes SCFNet, a modality-independent and domain-generalizable medical image segmentation method. This method introduces Dynamically Focused Convolution (DFC) and Interactive Attention Fusion Decoding (IAFD) modules. Experiments on 10 datasets covering four modalities demonstrate that SCFNet exhibits superior performance in both segmentation accuracy and generalization ability compared to 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.Inspired by the dynamic focusing mechanism of human vision, a novel network design incorporates two innovative and efficient modules: DFC dynamically adjusts convolutional weights and receptive fields to achieve adaptive feature extraction; IAFd utilizes wavelet transform and cross-layer interactive attention mechanisms to improve decoding and fusion efficiency. 2.Strong domain generalization and modality independence effectively address the critical domain offset problem in clinical medical image segmentation. 3.Comprehensive and rigorous evaluations were conducted on 10 datasets covering 4 imaging modalities. Both quantitative and qualitative results clearly demonstrate that this method outperforms 9 state-of-the-art methods in both seen and unseen domains. 4.This architecture, based on U-Net, is easily reproducible and can be integrated into existing medical image segmentation workflows. 5.Provides reliable and stable segmentation under cross-modality and cross-domain shifts.

  • 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.It is recommended to add ablation experiments to verify the effectiveness of DFC and IAFD respectively. 2.It is recommended to provide model efficiency analysis, including parameters, floating-point operations (FLOPs), and inference speed. 3.Test on a more challenging domain-transformed dataset to further verify generalization ability. 4.It is recommended to strengthen the description of the relevant work to better highlight the innovations compared to existing technologies. 5.It is recommended to clarify the implementation details of the DFC scaling settings to ensure the reproducibility of the results.

  • 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.

    (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?

    This paper proposes an effective and domain-general medical image segmentation method with significant performance improvements. However, this method lacks ablation experiments and efficiency analysis, and its innovation is limited in some aspects. Therefore, I suggest a weak acceptance.

  • 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.

    I think this paper could be acceptable due to its response.



Review #3

  • Please describe the contribution of the paper

    This paper proposes SCFNet, a modality-agnostic medical image segmentation model aimed at improving domain generalization across different clinical datasets and imaging modalities. The main contribution is a U-shaped segmentation architecture built around two proposed modules: Dynamic Focusing Convolution (DFC), which adaptively adjusts convolutional aggregation weights and receptive-field scales based on the input, and Interactive Attention Fusion Decoding (IAFD), which uses wavelet-based decomposition and cross-level interactive attention to improve decoder-side feature fusion. The paper evaluates the method on 10 datasets across dermoscopy, radiology, ultrasound, and colonoscopy, and reports improved performance in both seen and unseen clinical settings.

  • 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 paper addresses an important and practical problem. Domain shift is a major barrier to deploying segmentation models in real clinical settings, and the paper directly targets this issue by evaluating on unseen datasets across multiple modalities. This is a meaningful and clinically relevant problem formulation. 2) The method is reasonably well motivated and architecturally coherent. The paper presents a clear intuition inspired by dynamic visual focusing and translates it into two concrete modules, DFC and IAFD. The framework diagram on page 3 helps explain how adaptive perception and decoder fusion are integrated within the overall network. 3) The empirical evaluation is broad. The experiments cover four imaging modalities and ten datasets, with both seen-domain and unseen-domain testing. This is broader than many medical segmentation papers that focus on only one modality or one dataset pair. The quantitative results in Tables 1 and 2 show consistent gains over several strong baselines. 4) The reported results on unseen datasets are strong. The paper’s most convincing evidence is the improvement under unseen clinical settings, where the gains over prior methods become larger in some cases. This supports the claim that the method is aimed at generalization rather than only in-domain optimization.

  • 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 methodological novelty is only moderate, especially for DFC. The paper explicitly states that the feature-focusing part of DFC. In other words, an important component of the proposed method is already built on top of CondConv-style dynamic convolution, which reduces the novelty of this module. The paper would be stronger if it more clearly articulated what is fundamentally new beyond extending dynamic kernel weighting and adding scale attention. 2) The paper is not sufficiently differentiated from very closely related prior DG work. The most relevant prior work appears to be MADGNet [22], titled “Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multi-scale attention.” This prior work is extremely close in both problem setting and high-level design language. Although the paper compares against MADGNet in Tables 1 and 2, the method section does not sharply explain what conceptual gap remains after[A], and why SCFNet should be seen as a substantial advance rather than another dynamic/frequency-attention variant. Similarly, the paper itself cites domain and content adaptive convolution based multi-source domain generalization [3], which also suggests that adaptive convolution for medical DG is already an active direction. 3) The evaluation protocol is broad, but the exact training setup is somewhat unclear. It is not fully clear whether the model is trained jointly across modalities, separately per modality, or with some shared protocol across datasets. The paper presents itself as “modality-agnostic,” but the experiments are organized by modality-specific datasets and tasks. This ambiguity makes it hard to judge how strong the claimed modality-agnostic generalization really is. 4) Important ablation studies are missing. Given that the main claim rests on two new modules, I would expect ablations such as: baseline U-Net vs +DFC vs +IAFD vs full model, effect of the feature-focusing and scale-focusing branches separately, effect of the wavelet decomposition in IAFD, computational cost and parameter overhead. These analyses are important for understanding where the performance gain comes from, but they are not presented in the current manuscript. 5) Baseline selection is somewhat mixed for a domain-generalization paper. While the paper compares against several strong segmentation models, many baselines are generic segmentation architectures rather than specialized DG methods. Since the central claim is about domain generalization, stronger comparison to more dedicated DG segmentation methods would make the evidence more convincing. [A] Nam J H, Syazwany N S, Kim S J, et al. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multi-scale attention[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024: 11480-11491.

  • Please rate the clarity and organization of this paper

    Good

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

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

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

    (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?

    This paper studies an important problem and reports strong empirical results across multiple datasets and modalities. The breadth of the evaluation is a real strength, and the unseen-domain performance is encouraging. I also think the proposed architecture is sensible and the paper is generally well presented. My main hesitation is that the paper does not yet make a sufficiently strong case for novelty and methodological distinctiveness. The DFC module is explicitly built as an extension of existing dynamic convolution ideas, and the closest prior work, especially MADGNet [22], is highly similar in both problem framing and design philosophy. In addition, the lack of ablations makes it difficult to understand how much each proposed component really contributes, and the exact training setup for the modality-agnostic claim is not as clear as it should be. For these reasons, I currently place the paper slightly below the acceptance threshold, though a strong rebuttal clarifying novelty and protocol could improve my assessment.

  • 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 rebuttal resolved my concerns and after reading the other reviewers’ comments, I suggest acceptance



Review #4

  • Please describe the contribution of the paper

    The paper addresses the challenge of Domain Generalization (DG) in medical image segmentation, specifically focusing on “modality-agnostic” scenarios where models must generalize across different imaging modalities (e.g., MRI to CT) and diverse scanners. The primary contributions include:Self-Coordinated Focusing Network, Dynamic Focusing Convolution,Interactive Attention Fusion Decoding

  • 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 transition from static perception to “input-adaptive, target-aware” feature extraction is well-justified and biologically inspired. The authors conducted comprehensive experiments across multiple benchmarks. SCFNet consistently outperforms state-of-the-art methods.

  • 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.

    A more detailed analysis of FLOPs and inference latency would be beneficial. The proposed architecture consists of several complex components.However, the manuscript fails to provide a comprehensive ablation study to isolate and quantify the contribution of each individual module.

  • 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?

    1.Method novelty, 2.Robustness validation and efficiency comparison of methods (Param & flops) For detailed information, please refer to “major weaknesses”.

  • 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.

    The page limit cannot be used as an excuse for not to show ablation experiments of the base components. Reasonable space adjustment is also crucial(such as a large number of comparisons of old methods and a large number of visualizations are secondary). As a complete academic paper, ablation for base components is indispensable. Verbal promises do not represent academic rigor.



Author Feedback

We sincerely thank all Reviewers and the Meta-reviewer for your time and constructive comments. Below, we clarify the issues raised by Reviewers. Novelty and differences from prior DG (Domain generalization) works. (Reviewers 3, Meta-Reviewer1) Thanks for your valuable comments. Compared with CondaConv [4], DFC is not a simple addition of scale attention, but an innovative optimization of the dynamic mechanism for lesion-characteristic domain shifts. The novelty and differences lie in the following 3 aspects: In the dynamic coding stage, DFC introduces spatial-distribution awareness to preserve region-level spatial information. In the dynamic operating range, it not only generates dynamic convolutional weights but also dynamically adjusts the aggregation range. In the dynamic granularity, it extends image-level dynamics to region-level dynamic perception. These designs enable more adaptive and effective feature extraction under domain shifts. Compared with MADGNet [22] and DADC [3], SCFNet handles domain shifts in a different way. MADGNet mainly relies on multi-scale convolutional branches and frequency attention, which apply the same operations to all inputs and thus remain essentially static. Although DADC introduces dynamic feature extraction, its dynamicity is mainly limited to image-level convolutional weight generation based on global pooling. In contrast, SCFNet performs spatial-distribution-guided region-level dynamic perception by dynamically adjusting both weights and aggregation ranges. Together with advantage-information interaction fusion, SCFNet can better handle characteristic shifts across images and among different regions within an image, leading to more accurate and generalizable segmentation. Experimental setting and baseline selection. (Reviewer 3) SCFNet is trained separately for each modality with the corresponding seen data: two merged seen datasets for colonoscopy and one seen dataset for each other modality. The trained model is then tested separately on each seen and unseen dataset of the same modality without additional adaptation, to evaluate segmentation performance and cross-center generalization. Here, “modality-agnostic” means that SCFNet is a unified segmentation framework without requiring modality-specific priors or designs. We will optimize the description in the paper to avoid ambiguity. The baselines were selected by considering both method-level perception characteristics and the task-level need for generalization. We highly appreciate the reviewer’s suggestion to strengthen comparisons with dedicated DG segmentation methods, and will conduct more systematic evaluations in future work. Ablation and efficiency. (Reviewers 2, 3, and 4, Meta-Reviewer1) We apologize that these analyses are not fully presented due to space limitations. During preparation of the paper, we conducted ablation and efficiency analyses on DFC, IAFD, the overall architecture, the balance among parameters, computation consumption and performance, and so on. These analyses demonstrate the effectiveness of the proposed modules, architecture, and hyperparameters. We sincerely hope to share them in more detail during the conference discussion. Hyperparameter and cross-domain validation. (Reviewers 1 and 2) We agree that these analyses are important for evaluating stability and applicability. Current experiments focus on intra-modality cross-center generalization, with hyperparameters set based on preliminary analyses. We will conduct sensitivity analysis and explore cross-imaging-domain validation in future work. DFC scaling, distinctiveness, experimental description, and Fig. 2.(Reviewers 1, 2 and 3) We sincerely thank the reviewers for their valuable suggestions regarding the method description, implementation details, and figure readability. We will optimize the descriptions of the scaling implementation of DFC, method distinctiveness, and experimental settings, and adjust the layout of Fig. 2 to improve readability.



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.

    The paper addresses an important problem with a broad, well-executed evaluation and demonstrates strong generalization performance. However, the moderate novelty, insufficient differentiation from MADGNet, missing ablations, and lack of efficiency analysis keep it from being a strong accept. A rebuttal that clarifies novelty relative to prior work and provides missing ablation/efficiency results could strengthen the case.

  • 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.

    After reading the rebuttal, I recommend ‘Accept’ to this paper.



Meta-review #2

  • 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.

    The rebuttal clarifies the experimental protocol and better differentiates SCFNet from CondConv, MADGNet, and DADC. However, the central concerns remain unresolved. The paper provides no concrete efficiency analysis despite added architectural complexity, and the clarified protocol shows intra-modality cross-center testing rather than stronger modality-agnostic generalization. The reported results are promising, but without evidence isolating the proposed components and their computational cost, the methodological claims are insufficiently supported.



Meta-review #3

  • 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.

    While the proposed modality-agnostic segmentation framework shows broad empirical results across multiple datasets, consensus post-rebuttal highlights that the manuscript lacks indispensable component-level ablation studies, which cannot be excused by page limits. Furthermore, critical concerns remain unresolved regarding the limited methodological novelty relative to closely related prior work, the lack of clinical efficiency profiling, and ambiguous training details.



back to top