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

Abstract

Deep learning models face challenges in medical image segmentation under limited annotations, particularly in pediatric brain tumor cases. Simply augmenting pediatric datasets with adult cases is suboptimal due to anatomical and developmental differences. We propose a two-stage meta-reweighting framework: a coarse stage that performs cross-domain binary segmentation for stable localization, and a fine stage that refines pediatric-specific subregions with gradient-alignment-based reweighting. To further improve pediatric relevance, we introduce a gradient-similarity–driven retrieval mechanism that selects representative pediatric meta-cases as a stable reference set for weight computation, avoiding reliance on noisy online samples. This design enables effective loss reweighting while maintaining computational efficiency. Experiments on BraTS-PEDs demonstrate that the proposed method significantly outperforms state-of-the-art reweighting approaches, achieving a 6.1% improvement in Dice score and a 20.84 mm reduction in HD95, highlighting its robustness for pediatric brain tumor segmentation in low-data settings. Code is available at https://github.com/abod-alfakih/MDR2.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2361_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/abod-alfakih/MDR2

Link to the Dataset(s)

BraTS 2021: https://arxiv.org/abs/2405.18368 BraTS-PEDs 2024: https://arxiv.org/abs/2407.08855

BibTex

@InProceedings{AlAbd_CoarsetoFine_MICCAI2026,
        author = { Al-Fakih, Abdulkhalek AND Choi, Kyu Sung AND Kim, Dong-Hyun AND Ryu, Kanghyun AND Al-masni, Mohammed A.},
        title = { { Coarse-to-Fine Meta-Reweighting with Dynamic Retrieval for Adult-to-Pediatric Domain Adaptation in Tumor Segmentation } },
        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

    The paper presents a method for pediatric glioma segmentation that leverages both adult and pediatric glioma images during training. The approach is trained in two stages: the first performs binary tumor segmentation, while the second focuses on pediatric tumor parsing. In the second stage, a set of representative pediatric cases is identified every five epochs to reweight the influence of training samples based on their similarity to these cases. The results demonstrate that, compared with the selected methods, the proposed approach achieves improved performance.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper is well motivated. Pediatric glioma is underrepresented while adult glioma is a highly study task. This contrast highlights the potential impact of leveraging knowledge from adult data to improve pediatric applications.
    • The proposed method is interesting. Finding the more representative samples could lead to improve optimization and help mitigate overfitting to the adult cohort.
    • The method description and the figures are clear, and the paper is easy to follow
    • The ablation results show that the independent elements of the proposed approach improve the performance and complement each other
  • 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.
    • Comparison with state-of-the-art. The method shows improved results compared with several methods for training with noisy data, data scarcity, or domain adaptation. However, there are no comparisons against methods specifically developed for pediatric glioma segmentation, such as those proposed in the BraTS-PED 2023 challenge [a], which uses the same dataset as this work. Including such comparisons would provide a more task-relevant benchmark and better contextualize the contribution.
    • Missing baseline with only pediatric images. The introduction explains how training with both populations will affect the performance on the underrepresented pediatric cohort. However, it is unclear how the method performs when it is only trained on pediatric data. The first row in Table 1 may correspond to this setting, but it could also be trained only on adult images. Clarifying this point is important for understanding the benefit of incorporating adult data. Moreover, when compared to the BraTS-PED 2023 challenge results, the reported performance looks substantially lower, even though those methods were trained using only pediatric data. While the results are not directly comparable since its different evaluation sets (official test vs. selected test subset), the difference with several of those methods, mainly in ET, is large enough to raise questions about the method’s performance against approaches designed for this task.
    • The paper claims that the proposed approach has “lower computational and memory overhead” but does not include quantitative comparisons to support this statement
    • The introduction says that there are only 99 labeled cases for pediatric glioma analysis suggesting that that is the total number of samples available, but that is only the number of cases in the version of the dataset used. The more recent BraTS-PET 2024 has 261 [b]. Also, section 2 says that BraTS 2024 is used but the number of patients match those of 2023.
    • The order of sections 2.3, 2.4 and 3 is strange. The explanations of the baselines experiments are before the experimental setup. Also, section 2.3 should be closer to or merged with section 3.2 since they are highly related.

    [a] Fathi Kazerooni, Anahita, et al. “BraTS-PEDs: Results of the Multi-Consortium International Pediatric Brain Tumor Segmentation Challenge 2023.” Machine Learning for Biomedical Imaging 3.June 2025 issue (2025): 72-87. [b] Kazerooni, Anahita Fathi, et al. “The brain tumor segmentation in pediatrics (BraTS-PEDs) challenge: focus on pediatrics (CBTN-CONNECT-DIPGR-ASNR-MICCAI BraTS-PEDs). “ arXiv preprint arXiv:2404.15009 (2024).

  • 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 paper addresses an important problem and presents an interesting approach. The method is easy to follow, and the results suggest improvements over several baselines. However, key aspects of the evaluation are missing. In particular, there are no comparisons with methods specifically designed for the same task and dataset. Also, without a clearly defined pediatric-only baseline, it is difficult to assess the benefit of adding adult data. This particular clarification is necessary considering the motivation of how training with both cohorts tend to overfit to adult populations.

  • 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 provided clarifications and the new comparison with state-of-the-art methods in BraTS-PED improve the paper’s readability and contextualize the results better in the pediatric glioma segmentation task.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a two-stage framework for adult-to-pediatric brain tumor segmentation under limited pediatric labels. In the first stage, the model performs binary tumor localization on mixed adult and pediatric data to learn coarse tumor extent. In the second stage, it refines pediatric tumor subregions using meta-reweighting, where a dynamically updated set of pediatric “meta-cases” is selected by gradient similarity and used to reweight training samples according to alignment with pediatric gradients.

  • 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 an important and clinically relevant problem. Pediatric brain tumor segmentation is clearly a low-data regime, and the manuscript is well motivated around the difficulty of leveraging adult data without simply inheriting adult-specific biases.

    • The proposed formulation is intuitive and reasonably well matched to the problem. The idea of separating coarse tumor localization from fine pediatric-specific subregion refinement is sensible, and the use of the coarse mask as an auxiliary channel in the fine stage is clearly described.

    • The dynamic retrieval mechanism is a potentially useful extension over fixed or arbitrary meta-sample selection. The method computes a mean pediatric gradient, scores pediatric cases by cosine similarity, and uses the top 10 as meta-cases for reweighting.

    • The empirical section includes more than a single benchmark comparison. The paper provides an ablation study, comparisons against several reweighting/domain adaptation baselines, four-fold results, and backbone sensitivity experiments.
    • Within the paper’s own experiments, the full method consistently improves over its internal baselines.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    • The paper has multiple reference-integrity problems. The manuscript states that it uses “BraTS 2021 (1,251 adult cases) [11],” but reference [11] in the bibliography is actually the 2024 BraTS post-treatment glioma challenge paper, not the standard BraTS 2021 adult pre-operative dataset. The baseline section also describes DA [6] as “supervised Domain Adaptation,” while the cited reference is explicitly titled Unsupervised domain adaptation for pediatric brain tumor segmentation. In addition, “L2RW [7]” is cited as a skin-lesion segmentation paper, whereas the standard L2RW reference is Ren et al., Learning to Reweight Examples for Robust Deep Learning (ICML 2018). I also note that MGR-DAS is described as [9] in the baseline section but appears as [10] in Table 2.Reference [1] is relevant, but poorly placed. It is a prior pediatric transfer-learning MICCAI paper, so it fits better as prior work on adult-to-pediatric adaptation than as support for the claim about tumor incidence.

    • Several of the strongest claims are overstated relative to the actual evidence. In particular, the abstract states that the method “significantly outperforms state-of-the-art reweighting approaches, achieving a 10.7% improvement in Dice score and a 24.6 mm reduction in HD95.” Those numbers correspond to the gain over naive mixed training in Table 1, not the strongest competing reweighting baseline in Table 2, so the abstract wording is stronger than the evidence clearly supports. Likewise, the paper claims “state-of-the-art pediatric segmentation performance,” but it does not compare against the leading methods summarized in the BraTS-PEDs challenge results paper, where top pediatric approaches included nnU-Net/Swin UNETR ensembles, Auto3DSeg, and self-supervised nnU-Net variants.

    • The evaluation protocol is ambiguous. Section 2.1 states that BraTS-PEDs is randomly divided into 70 training and 29 testing cases, but in the same paragraph it says the experiments are repeated under a 4-fold cross-validation scheme. Section 2.4 again refers to robustness evaluation using four-fold cross-validation. It is therefore unclear whether the main reported results are based on a fixed 70/29 split, cross-validation averages, or a mixture of both. In a 99-case pediatric dataset, this ambiguity is a major issue.

    • Reproducibility is limited by missing implementation detail. The paper gives the high-level equations and some training settings, but several important details are absent: how sample-level gradients are computed in practice for 3D segmentation, whether gradients are computed over all parameters or a subset, how patch-level gradients are aggregated, how checkpoints are selected, and how hyperparameters such as retrieving 10 meta-cases every 5 epochs were chosen.

  • 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

    Thank you for addressing an important pediatric imaging problem. The paper’s central idea is reasonable and the component ablations are helpful. My main recommendation is to improve the paper’s precision: audit the references carefully, clarify the exact evaluation protocol, and provide a more complete description of the implementation details needed for reproducibility. I also encourage the authors to calibrate a few of the strongest claims more conservatively to better match the presented evidence.

  • 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?

    I rate this paper as 3 (Weak Reject) because it presents a reasonable and potentially useful idea, but the current submission falls slightly below the acceptance threshold due to concerns about reference accuracy, evaluation clarity, and reproducibility. The core method is a two-stage adult-to-pediatric segmentation framework combining coarse binary localization, fine pediatric subregion refinement, and dynamic gradient-similarity retrieval for selecting pediatric meta-cases. This is a coherent approach, and the paper shows consistent improvements over its internal baselines.

    The main factors lowering my score are as follows. First, several references are misused or misdescribed, which reduces confidence in the technical framing and literature review. Second, the evaluation protocol is not clearly specified, making it unclear which setup underlies the main reported results and how model selection was performed. Third, some claims are stronger than the evidence supports, especially the statement of “state-of-the-art pediatric segmentation performance,” given the limited comparison set and the lack of comparison to stronger pediatric or challenge-leading methods. Finally, reproducibility is only moderate in the current draft: while the high-level method is understandable, important implementation details are missing, particularly regarding sample-level gradient computation in 3D segmentation, the practical implementation of dynamic retrieval, and the basis for the efficiency claims. The statement that code will be released upon acceptance is positive, but the manuscript alone is not yet detailed enough for confident reproduction.

    Overall, I view this as a promising but not yet fully convincing submission. The problem is important, the method is plausible, and the empirical trend is encouraging, but the paper would need more careful citation handling, a clearer and unambiguous evaluation protocol, and more precise reporting of implementation details and training procedure to reach the acceptance bar.

  • 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.

    This remains a relatively close decision for me, but the rebuttal addresses enough of my concerns that I am willing to support Accept, provided that the authors follow through on their stated commitment to release the code and to correct the citation/reference issues in the final version. The clarification of the evaluation protocol resolves my main ambiguity about how to interpret the results, and the added runtime/implementation details and stronger-pipeline experiment improve both the reproducibility and the empirical positioning of the work. While I still see some weaknesses, I think the paper is now just above the acceptance threshold, especially if the promised code release and reference corrections are carried through in the camera-ready version.



Review #3

  • Please describe the contribution of the paper

    This paper addresses the critical problem of pediatric brain tumor segmentation under limited annotated data by leveraging abundant adult data. The authors propose a two-stage coarse-to-fine meta-reweighting framework with dynamic gradient-similarity retrieval to mitigate the adult–pediatric domain gap. The coarse stage performs cross-domain binary tumor localization, while the fine stage refines pediatric-specific subregions using gradient-aligned sample reweighting. Experiments on BraTS-PEDs show clear improvements over state-of-the-art reweighting and domain adaptation methods in both DSC and HD95.

  • 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.Well-designed two-stage strategy: Separates cross-domain consistent localization and pediatric-specific refinement, which is intuitive and effective. 2.Efficient reweighting: Avoids costly bi-level optimization and maintains low computation overhead, suitable for clinical deployment. 3.Clear writing and logical structure: The paper is well organized, with clear figures and formulas.

  • 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.How were the meta-cases selected, what is the clinical rationale, do they include rare cases? 2.Lack of generalization analysis in other pedritric datasets

  • 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?

    The method is technically sound, well-motivated, and clearly presented.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Accept



Author Feedback

We thank the reviewers for their insightful feedback and positive assessment of our coarse-to-fine meta-reweighting framework. We address the concerns below: (1) Comparison with BraTS-PED Challenge Winners (R1, R2): The top-performing methods in the BraTS-PED 2023 challenge, such as CNMCPMI2023 and Auto3DSeg, rely on ensemble pipelines that combine U-Net, Swin-UNETR, and ResNet architectures with larger model capacities and patch sizes. In contrast, our experiments use single U-Net or Swin-UNETR models with a smaller 64³ patch size and lower-parameter models due to limited computational resources, which may particularly affect ET performance in pediatric-only training. Nevertheless, as shown in Table 3(B), the proposed method consistently improves both backbones, indicating that it functions as a plug-and-play, backbone-agnostic learning framework. To directly address the reviewers’ concerns, we integrated the proposed method into the challenge-winning CNMCPMI2023 pipeline ( U-Net and Swin-UNETR with additional post-processing) on our split. This improved the average Dice score of the CNMCPMI2023 pipeline from 0.763 to 0.822, further demonstrating compatibility with SOTA ensemble pipelines. (2) Baseline Clarification (R1): The first row in Table 1 is indeed the pediatric-only baseline. We will clarify this explicitly in Table 1 caption. (3) Cross-Validation Ambiguity (R2): We would like to clarify that all main results are based on a fixed 70/29 pediatric train/test split for fair comparison with prior work. The additional 4-fold cross-validation experiments (25, 25, 25, and 24 testing cases) were conducted separately as a robustness analysis (only Table 3(A)). We agree Sec. 2.4 could be clearer, and we will revise it to avoid ambiguity. (4) Reference Inconsistencies (R2): We sincerely thank R2 for the meticulous audit. DA [6] is primarily an unsupervised domain adaptation (UDA) study; however, the paper also reports all supervised DA results as an upper-bound reference. In our comparison, we selected the best-performing supervised DA from this paper. For L2RW [7], we cited the medical-image implementation paper (skin lesion) that applied L2RW, but we agree that the original ICML 2018 L2RW paper should also be cited. References [9] and [10] are conference/journal versions of the same work, and we will fix the numbering inconsistency, and the BraTS 2021 [11] citation will be corrected accordingly. (5) Reproducibility, Implementation Details, and Computational Efficiency (R1, R2, R3): Parameter-Efficiency: To minimize overhead, our method avoids bi-level optimization and restricts gradient computation to only over the encoder and decoder (∼4.45M parameters), excluding the ViT layers (∼100M parameters). As a result, our method requires 3.26 min/epoch, compared to 26.42 (MWN), 28.32 (L2RW), and 6.82 (MGR-DAS). We will add it to Table 2 Gradient Computation: We use micro-batching to compute sample-level gradients within 3D patches. Full implementation details will be available on GitHub upon acceptance, per MICCAI’s double-blind policy. (6) Meta-case Selection & Generalization (R3): Meta-cases are selected based on similarity to the average pediatric gradient G_ref (section 2.2 (i)), which prioritizes representative pediatric tumor patterns (e.g., missing/small edema regions) while reducing the influence of rare cases (e.g., large ED regions). Clinically, this helps the model focus on consistent pediatric characteristics rather than adult-dominant patterns. For testing on another dataset, we tested our method on pediatric brain tissue segmentation, and due to space limits, included only pediatric tumor segmentation. (7) Paper Organization (R1,R2): We agree that the organization of Secs. 2.3, 2.4, and 3 can be improved for readability. In addition, the reported 10.7% Dice improvement is relative to naive mixed training rather than the strongest competing (MGR-DAS), which is 8.5%. We will revise them in the revised version.



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 presents an interesting two-stage meta-reweighting method for adult-to-pediatric domain adaptation under limited pediatric segmentation annotations. However, the paper is unclear and not rigorously written, leaving several parts difficult to interpret. The major issues are: (1) the evaluation protocol is unclear (R1R2), it is uncertain whether the main results are from cross-validation or a separate test set. (2) The paper also lacks comparison with BraTS-PED challenge-specific SOTA baselines, making it difficult to assess the actual contribution of the proposed domain adaptation (R1R2). More importantly, (3) the paper contains multiple citation errors, including issues with cited baselines and datasets (R1R2). (4) overstated claims without sufficient numerical evidence, such as “lower computational and memory overhead” (R1) and “state-of-the-art pediatric segmentation performance” (R2). Overall, the method is interesting, but the paper is not sufficiently rigorous and the rebuttal should address these concerns from the reviewers.

  • 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 addressed the reviewers’ major concerns, and all reviewers now recommend acceptance. To further improve the paper, the authors are encouraged to thoroughly correct inconsistencies in references and tables, including those not mentioned, and to commit to releasing the code upon acceptance. In addition, the BraTS-PED dataset has evolved over time, and updated versions with much more subjects are now available.



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.

    After considering the rebuttal and updated reviews, I recommend acceptance. All three reviewers support accepting the paper, and the rebuttal has sufficiently addressed the main concerns. Although some weaknesses remain, the paper is above the acceptance threshold, provided that the authors follow through on the promised code release and corrections in the final version.



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.

    Following the rebuttal period, all reviewers recommended acceptance. Most acknowledged that the paper addresses an important problem using a reasonably novel approach. The authors should pay close attention to writing quality and formatting in the camera-ready version.



back to top