Abstract

Precise biomedical image segmentation is crucial for clini- cal diagnosis. Geometric cues (e.g., boundary, shape, and topology) can improve structural consistency, yet most are task-specific and lack a unified geometric foundation that generalizes across organs and modalities. We are motivated by the observation that many medical segmentation targets exhibit globally near-convex shapes. A convex region is one in which any two interior points can be connected by a line segment entirely contained within the region. In practice, medical targets may exhibit small local concavities or boundary irregularities; we refer to such globally convex-like shapes as near-convex. Motivated by this, we derive Hadwiger Shape Priors from Hadwiger’s theorem as an interpretable global regularizer using three 2D measures: area A, perimeter P , and Euler characteristic χ, enabling transfer across organs and modalities. However, because medical datasets are shape-heterogeneous, enforcing near-convex priors uniformly can over-regularize non-convex anatomy (with significant concavities), washing out concavities and fine details and degrading segmentation accuracy. To address this challenge, we propose Conflict-Aware Objective Balancing (CAOB), which integrates shape priors with segmentation in a gradient-aware manner. For each prior, CAOB removes only the gradient component that conflicts with segmentation while preserving the remaining aligned component, and adaptively regulates objective influences to prevent prior dominance. This enables stable use of shape priors on shape-heterogeneous data without erasing genuine concavities or fine structural details. We call this plug-and-play framework HadBalance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0859_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/NatsuGao7/HadBalance

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GaoZhu_HadBalance_MICCAI2026,
        author = { Gao, Zhuangzhi AND Zhou, Feixiang AND Zhao, He AND Chen, Wenhan AND Luo, Ruiyu AND Wang, Xin AND Qin, Hongyi AND Wu, Zhongli AND Meng, Yanda AND Zhao, Yitian AND Shantsila, Alena AND Lip, Gregory Y. H. AND Shantsila, Eduard AND Zheng, Yalin},
        title = { { HadBalance: A Plug-and-Play Unified Global Geometric Prior Framework for Generalizable Biomedical 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 introduces a Hadwiger-based prior for training segmentation models in medical imaging. Based on the observation that most anatomical structures are approximately convex, the authors use geometric properties (area, perimeter, and Euler characteristic) which, according to Hadwiger’s theorem, characterize convex bodies. To prevent the prior from over-regularizing the model during training, they propose a COAB module. This module suppresses gradients from the prior when they conflict with the segmentation loss, ensuring that the prior does not dominate the learning process.

  • 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 manuscript is generally well-structured and clearly written. It provides a detailed explanation of the Hadwiger prior and the underlying theorem, making the work accessible and understandable for readers
    • The proposed method is well-formalized and presented in a way that is straightforward for the reader to follow
    • The use of real-world datasets, although limited to three, is a positive aspect, as it allows readers to evaluate the applicability of the method in practical scenarios
    • The authors have made the source code publicly available. While the training and inference scripts are not included, the provided code still enables readers to implement the method with minimal effort
    • The figures are carefully curated and, when viewed on a PC with the ability to zoom, are clear and easy to interpret
  • 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 weaknesses:

    • The authors state that they “observe that most medical segmentation targets are globally near-convex”. I think this is a very strong claim which is not supported by any evidence in the paper. It is also informal, how is it measured or quantified in this context? It is also unclear which types of medical structures were considered when making this statement. Even if it was an assumption, this is quite strong and should be better justified, either through empirical analysis or references to prior work

    • The authors said in the manuscript “Incorporating geometric cues [. . . ] is a common way to improve structural consistency; however geometric priors are largely task-specific [. . . ] and lack a unified foundation that generalizes across most organs and modalities and supports transferable geometric constraints”; the term “foundation” in this sentence feels somewhat vague and could be replaced with a more precise description of what is lacking. Also, the statement appears to frame task-specific geometric priors as a limitation due to their lack of generalizability, but, in medical imaging it is often both reasonable and expected for priors to be tailored to specific anatomical structures, as different organs exhibit distinct shapes and variability. The sentence needs to be rephrased for clarity and precision

    • The AGB module, as much as it is well formalised and if I understood correctly, computes alpha(t) for each optimisation step, using the non-linear least-squares solver from scipy. This introduces a number of significant practical issues that are not discussed anywhere in the paper (except for a brief two word mention in the conclusion). First of all, the complexity of the scipy solver scales with the number of network parameters at least linearly, and more importantly it is embedded within another optimization process, so it adds a lot of computational overhead. Moreover, scipy does not support GPU computation; therefore, if the authors indeed used the scipy solver, it implies that either data must be transferred between GPU and CPU during each optimization step or the model training is being done on CPU, both of which are highly inefficient. The drawback of the training time is not discussed, and comparison of training costs are not reported in the experiment and result section

    • The authors do not compare their method with other state-of-the-art approaches on each dataset. As a result, it is not possible to assess whether their work achieves comparable performance to existing methods on the datasets they use Standard deviations and COnfidence intervals are missing in every table of quantitative results. Given also the fact that train/validation/test set sizes are not reported, it is difficult to assess whether their method provide statistically significant results from the other methods, given that performance differences are very low for the most part (e. g. 86,19 vs 86,35)

    • In the conclusions, the authors state: “In this paper, we observe that many medical segmentation targets exhibit globally near-convex geometry”. This claim is too strong and not supported by the evidence presented. The manuscript only includes a few examples from three medical targets, which is insufficient to generalize broadly. Additionally, the statement is informal and should be either supported with empirical analysis, clarified as an assumption, or softened to reflect the limited scope of the evidence

    Minor weaknesses:

    • The images in the manuscript are quite small, making it difficult for readers, especially in printed format, to distinguish details such as text and formulas.

    • In the introduction, where the authors outline their key contributions, colons are overused: “Our key contributions are as follows: 1: [. . . ]”. I recommend using a closing parenthesis after the item number instead, like “1) […]”, to improve readability

    • The datasets used in the experiments are barely described, consisting of just a three-sentence paragraph with only citations. There is no discussion of what the datasets contain, the number of observations, or how the data were collected

    • In the experimental setting, the authors train several backbones, they measure their performance, and perform a new training from scratch using the strongest backbone with their COAB module (for each dataset). While this comparison between backbone A + COAB versus backbone A alone is reasonable, it is unclear why only the strongest backbone was selected for this experiment. The manuscript does not provide any justification for this choice, why not add COAB to all the backbones for a more complete comparison?

    • DRIONS-DB dataset is missing in the ablation study, with no justification in the text; also the Module / Loss columns are repeated in the table

    • “HadBalance outperforms NASH and Fairgrad” is very strong, given the fact that only in few comparisons there is a clear separation between the methods (3 out of 9). The claim needs to be softened

    • The authors perform a comparison of their method against two gradient-optimisation methods (NASH and FairGrad), however they do not explain in a few words what they are and how they operate, for more clarity they should be explained briefly

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    The manuscript does not include comparisons with state-of-the-art methods on the datasets used, making it difficult to assess the relative performance of the proposed approach. Additionally, the quantitative results lack standard deviations or confidence intervals, and the reported differences in performance are very small, which limits confidence in the findings. The manuscript also does not discuss any limitations of the study,

  • 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 article proposes:

    • “Hadwiger Shape Priors” adapted to 2D medical image segmentation and differentiable surrogates for them;
    • a framework to integrate such regularization priors into the loss function used to train neural networks for segmentation tasks.
  • 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 regularization framework does not need hand-tuned weights for the different components of the loss.
    • A detailed description and justification for the different “modules” of the regularization framework is provided.
    • Experiments have been conducted on different datasets and with different network architectures.
    • Authors have conducted an ablation study to assess the contribution of their different modules and their potential synergy.
    • Some source code is provided.
  • 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:

    • No statistical analysis of the results, notably regarding the significance of the differences observed.

    Minor:

    • The 3D segmentation case is not mentioned at all. Can the proposed shape priors be generalised to 3D?
    • In introduction, page 2, paragraph 2 [ref needed], the authors claim that “most medical segmentation targets are globally near-convex” without any reference or experiment backing this claim.
    • The is no mention of the impact of the regularization on the computer power requirements (how much longer does it take? Does it require more VRAM?)
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    Since experiments use open datasets and open source frameworks such as nnUNet, it would have been nice to provide exhaustive-enough code to reproduce the experiments.

  • 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?

    Topological constraints are interesting for the MICCAI community and the constraints derivation and gradient management for the loss is a mathematically sound way to introduce those that can be extended to other applications, even if the numerical results are not very impressive compared to the baselines.

  • 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.

    even if generalisation to 3D is not straightforward and numerical improvement is not impressive, it is an interesting idea



Review #3

  • Please describe the contribution of the paper

    This paper proposes a plug-and-play biomedical segmentation framework called HadBalance, which employs a unified global geometric prior and conflict-aware optimization algorithm. It introduces Hadwiger shape priors for near-convex shape regularization and designs the CAOB algorithm to adaptively resolve gradient conflicts. Experimental results show that the framework achieves consistent performance improvements across different organs and modalities.

  • 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.This method uses area, perimeter, and Euler characteristic as a theoretically sound, interpretable, and transferable global shape prior for medical segmentation. Unlike task-specific geometric constraints, this prior generalizes to different organs and modalities, representing a significant conceptual advancement. 2.This method addresses a key practical problem: shape priors often over-regularize non-convex anatomy. CAOB automatically removes conflicting gradient components (PGPs) and adaptively balances the loss (AGB) without requiring manual weight adjustments, making it stable and easily integrated into existing models. 3.The framework was validated on three different datasets (carotid artery, polyp, optic disc) using six mainstream backbone networks. Its performance outperforms strong baselines including nnUNet, TransUNet, and state-of-the-art multi-task optimizers (Nash-MTL, FairGrad), particularly showing consistent gains in boundary accuracy. 4.HadBalance is model-agnostic and can be directly integrated into existing segmentation pipelines, demonstrating strong clinical and engineering potential.

  • 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.CAOB requires separate backpropagation for each loss term, which increases training time and GPU memory usage compared to standard single backpropagation methods. This paper does not provide efficiency analysis or optimization strategies. 2.Experiments were conducted only on three near-convex-dominated 2D datasets. Its performance on highly non-convex organs (e.g., liver, lung, heart) and 3D volume data has not been tested, limiting its generalization potential. 3.This paper does not compare with state-of-the-art learnable shape priors or topology-aware segmentation methods, making it difficult to determine the relative advantages of Hadwiger constraint-based methods. 4.Convergence of fixed-point iteration in AGB is not analyzed. Stability and sensitivity to hyperparameters (e.g., pooling size, ρ) have not been thoroughly discussed.

  • 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 a novel and effective segmentation framework, which is well-motivated and yields significant experimental results. However, the framework has limitations in terms of computational cost, validation scope, and theoretical analysis. Despite its solid core idea and well-functioning method, I recommend 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.

    Methodology seems simple, but the topic has its meaning, so I think this paper could be acceptable



Author Feedback

We thank the reviewers and ACs. Responses are as follows (R: Reviewer, Q: Question): The near-convex assumption: We agree that “most medical segmentation targets are globally near-convex” was too strong; our intent was to use near-convexity as a qualitative motivation, not to make a universal empirical claim over all medical structures. We will revise “most” to “many”, soften “we observe that”, and add convex/star-convex prior references (Liu et al., 2020; Liu et al., JMIV 2025; Royer et al., CVPR 2016; StarDist, MICCAI 2018) with examples such as optic disc, nuclei, and cardiac structures. HadBalance does not pre-classify samples as near-/non-convex; instead, it applies A/P/χ priors to all samples and uses CAOB to suppress conflicting prior-gradient components when these priors would over-regularize genuine non-convex structures [R1Q1Q5, R2Q3]. The missing benchmarks: Dataset-specific SOTA methods such as Polyp-PVT (Dong et al., 2023) and Meng et al. (MICCAI 2022) are useful per-dataset baselines, but our experiments test HadBalance as a plug-and-play global geometric-prior framework rather than a separately tuned SOTA pipeline for each dataset [R1Q4]. Shape-prior methods such as You et al. (TMI 2025) and Song et al. (IJCAI 2024) introduce architecture-level modules/statistical shape models, so a direct comparison would mix architectural and training effects with our Hadwiger-prior and CAOB contributions. Topology-aware methods such as Hu et al. (NeurIPS 2019) and Li et al. (TMI 2025) mainly target topology preservation for topology-critical structures, such as vessels, and address a different objective from our work [R4Q3]. For [R1Q9], we selected the strongest backbone on each dataset as a stricter setting to test whether HadBalance can further improve an already strong baseline. The efficiency analysis: CAOB follows the per-objective-gradient paradigm: separate gradients are needed to estimate conflicts and balance objectives. The A/P/χ regularizers are lightweight pooling/convolution/gradient-magnitude operations; the main overhead is multiple backward passes, which is shared by methods such as FairGrad and Nash-MTL. The SciPy solver is not applied to the full gradients or network parameters. After per-objective gradients, we compute pairwise inner products on GPU to form a small K×K Gram matrix (K=4), and SciPy only solves the resulting 4-dimensional α problem. Thus, the CPU transfer is limited to this small Gram matrix, not full gradients or parameters, and training remains GPU-based [R1Q3, R2Q4, R4Q1Q4]. Regarding convergence, AGB does not perform high-dimensional fixed-point iteration; it solves a low-dimensional constrained least-squares problem for K=4 weights, and stability is empirically supported by CAOB ablations and gradient-alignment analysis. Sensitivity was under-discussed: D = {7,15,31} was fixed across experiments for multi-scale A/P/χ maps, while ρ is a scale-control term whose absolute effect is mitigated by the subsequent α normalization rather than acting as a dataset-specific loss weight [R4Q4]. Generalization analysis: 3D extension is feasible, but not by directly reusing the 2D A/P/χ formulation; it would require 3D intrinsic volumes and a redesigned differentiable 3D χ computation [R2Q2, R4Q2]. Highly non-convex organs are outside the validated scope of this submission, since Hadwiger priors are primarily motivated by near-convex geometry, and we therefore do not claim generalization to such organs here [R4Q2]. Results Analysis: We agree that the CUBS DSC margin is small and should not be judged from a single Dice value alone. Full HadBalance improves the strongest CUBS backbone by +1.21 BFScore, and on the more shape-heterogeneous CVC-ClinicDB, it improves over LS by +1.04 DSC, +1.50 IoU, and +1.64 BFScore, indicating that CAOB more robustly integrates Hadwiger priors without manual loss-weight tuning [R1Q4]. We will add confidence intervals to better reflect statistical uncertainty [R1Q4, R2Q1].




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 idea is not completely obvious, and the theoretical foundation is interesting, but whether it will have real impact in medical AI is still unclear because the evidence is limited. The three reviewers gave recommendations of reject, accept, and weak accept; I’d like the authors to partake in rebuttal (issues: the near convex assumption, the missing benchmarks, the efficiency analysis, etc.).

  • 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.

    Despite a split from reviewers, I recommend acceptance. I think the CAOB mechanism addresses the practical over-regularization problem in a mathematically sound way. The rebuttal adequately resolved the two sharpest concerns: the near-convex assumption was conceded and softened with prior-work support, and the SciPy efficiency objection was clarified. Missing benchmarks against architecture-level shape-prior methods and modest performance margins remain a limitation, but are defensible given the plug-and-play framing and the meaningful boundary-level predictions. I encourage authors to amend their camera-ready with proper discussions.



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.

    I recommend rejection. The paper addresses an interesting direction—using global geometric priors for biomedical segmentation—but the current submission does not sufficiently support its claims of being unified, generalizable, and plug-and-play.

    The main concern is that the core assumption, namely that many biomedical targets are globally near-convex, is not rigorously justified. The connection between Hadwiger’s theorem and the proposed practical losses is also overstated: area, perimeter, and Euler characteristic do not uniquely encode near-convexity and may be inadequate for complex or highly non-convex anatomy.

    The methodological novelty is limited. The proposed regularizers resemble existing shape/topology losses, and the conflict-aware optimization component is not clearly distinguished from prior multi-objective gradient-balancing methods. The derivation and practical stability of the adaptive weighting strategy are also insufficiently analyzed.

    The experiments are too narrow for the paper’s broad claims. Evaluation is limited to three relatively small 2D datasets, with no strong evidence on more challenging non-convex, thin-structure, multi-component, 3D, or cross-domain settings. The reported gains are mostly modest, and statistical validation, runtime analysis, and stronger comparisons to recent shape/topology-aware segmentation methods are missing.



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.

    After reviewing the authors’ rebuttal and the post-rebuttal recommendations from reviewers, I recommend Accept. The rebuttal effectively addresses the key concerns.



back to top