List of Papers Browse by Subject Areas Author List
Abstract
In volumetric medical image segmentation, semi-supervised learning (SSL) leverages unlabeled data to reduce the reliance on scarce expert annotations. However, standard optimizers treat gradients from labeled and unlabeled objectives uniformly, despite their differing reliability. We propose Hierarchical Cautious Optimization (HCO), an optimizer-level method that enforces a trust hierarchy between gradient sources. HCO computes momentum using gradients from labeled data and incorporates unlabeled gradients only when their update direction is aligned with this labeled-driven momentum. The method integrates seamlessly into existing momentum-based optimizers with negligible computational overhead. Across three datasets, HCO consistently improves performance, with notable gains on a challenging fetal MRI dataset where Dice scores for fetal lungs and liver increased from 0.68 to 0.84 (+24\%) and from 0.71 to 0.82 (+16\%), respectively. These gains across optimizers and datasets show that HCO is a practical enhancement for SSL segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0945_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{OleAar_Hierarchical_MICCAI2026,
author = { Olender, Aaron AND Rabinowich, Aviad AND Khawaja, Jayan AND Ben Bashat, Dafna AND Joskowicz, Leo},
title = { { Hierarchical Cautious Optimization for Semi-supervised Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper propose Hierarchical Cautious Optimization (HCO), an optimizer-level method that enforces a trust hierarchy between gradient from labeled and unlabeled objectives. HCO computes momentum using gradients from labeled data and incorporates unlabeled gradients only when their update direction is aligned with this labeled-driven momentum. It can be integrated seamlessly into existing momentum-based optimizers with negligible computational overhead.
- 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 proposed formulations are straightforward, intuitive, and clearly presented, making them easy to follow. The evaluation compares the proposed methods against multiple existing momentum-based optimizers across three different datasets. Extensive experiments are conducted to validate the effectiveness and generalizability of the proposed method.
- 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 principle of HCO is straightforward and somewhat innovative, but its convergence relies on a mild alignment condition. A more in-depth theoretical analysis would be helpful. Relying solely on gradient direction alignment may compromise the diversity of object predictions, particularly when considering the abundance of unlabeled images and the wide variation in object shapes and sizes. Only a single semi-supervised learning framework (i. e. , BCP) is used in this study.
- 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?
It would be better to (1) employ more semi-supervised segmentation frameworks (not merely BCP) and more network backbones for effectiveness and generalizability assessment; (2) provide more in-depth theoretical analysis for the proposed HCO; (3) extend gradient direction alignment beyond plain binary masking.
- 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 believe the authors’ rebuttal is acceptable, and the paper can be published in the MICCAI conference.
Review #2
- Please describe the contribution of the paper
The paper proposes a simple optimizer-level modification for semi-supervised learning. The main idea is to treat labeled and unlabeled gradients differently by introducing a trust hierarchy: the update direction is driven by labeled gradients, while unlabeled gradients are only used when they are aligned with that direction.
This results in a plug-and-play variant of standard momentum-based optimizers that aims to reduce the negative impact of noisy pseudo-labels. The method is evaluated on several medical image 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.
1.The paper points out a reasonable issue in semi-supervised training: gradients from labeled and unlabeled data are not equally reliable, but standard optimizers treat them the same. This is a valid observation, especially in medical imaging where pseudo-label noise can be significant.
2.The method itself is simple and easy to implement. It works at the optimizer level and does not require changing the model, loss, or training pipeline, which makes it easy to try in practice. The idea is fairly general and can be applied to different optimizers such as SGD and Adam. The additional computational cost is also minimal.
3.The experiments cover multiple datasets and settings, which helps demonstrate that the method is not tied to a specific task or framework.
- 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 main idea is quite simple and somewhat heuristic. The method essentially relies on an element-wise alignment check between labeled and unlabeled gradients, and it is not entirely clear why this specific design is the most effective choice compared to more principled alternatives (e.g., soft weighting, uncertainty-based filtering, or gradient projection methods). 2.The theoretical analysis is limited. The convergence result depends on an additional alignment assumption, but this assumption is not well justified and seems difficult to guarantee in practice, especially given that the paper itself shows that gradient conflicts are common. 3.The reported improvements, particularly on the fetal MRI dataset, are unusually large for an optimizer-level modification. This raises some concerns about the experimental setup. It would be helpful to see more thorough validation (e.g., more seeds, clearer data splits, or additional benchmarks) to better support these results. 4.The experimental evaluation could be stronger. The paper mainly builds on a small number of SSL baselines and does not include comparisons with more recent or stronger methods. In addition, there is no comparison with existing approaches that explicitly address gradient conflicts, which seems closely related to the proposed idea. 5.Some experimental details are not entirely clear (e.g., consistency of evaluation protocols across datasets, number of runs, and data splitting strategy), which makes it harder to assess the robustness 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 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.
(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 a relevant issue in semi-supervised learning by considering the different reliability of labeled and unlabeled gradients, and proposes a simple optimizer-level modification. The idea is intuitive and easy to implement, and the plug-and-play nature is a practical advantage.
However, the overall contribution is somewhat limited. The proposed method mainly relies on a simple alignment-based gating mechanism, and it is not entirely clear why this specific design is preferable to other existing approaches for handling noisy or conflicting gradients. The theoretical analysis is also relatively weak, as it depends on an additional assumption that is not well justified.
In addition, the experimental results raise some concerns. The reported improvements, especially on certain datasets, are unusually large for an optimizer-level change. The evaluation could be strengthened with more consistent experimental protocols, stronger baselines, and more thorough validation (e.g., additional runs or datasets).
Overall, while the paper presents a reasonable and practical idea, the current version does not yet provide sufficiently strong evidence or methodological depth to clearly meet the acceptance bar.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper improves semi-supervised segmentation from an interesting aspect, i.e., the gradient. Building on the motivation that gradients from labeled and unlabeled data have different reliabilities and should be treated differently, this paper proposes the hierarchical cautious optimization, which maintains the momentum of gradients from labeled data and uses the direction of labeled-driven momentum to filter the unlabeled gradients. Experimental results on three datasets with various semi-supervised methods and optimizers demonstrate the effectiveness of the proposed method.
- 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.Innovative aspect to improve semi-supervised segmentation. This paper improves the performance of semi-supervised segmentation from the aspect of gradients’ reliability, which is reasonable and often ignored in existing works. 2.The experimental validation is comprehensive. Results on different semi-supervised methods and optimizers demonstrate the efficacy of the proposed method.
- 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.Lack of comparisons on SOTA semi-supervised methods. While the authors validate their approach on two existing semi-supervised frameworks (BCP and CMT), both of these baselines were introduced in 2023.To better position the contribution and evaluate its general applicability, it is important to assess whether the proposed method remains effective when integrated with more recent and advanced techniques, such as UniMatch V2 [R1] (TPAMI 2025). Including such comparisons would strengthen the paper’s relevance and demonstrate the robustness of the method in contemporary research contexts.
2.The proposed hierarchical cautious optimization method heavily relies on gradient momentum derived from labeled data. This reliance raises a particular concern in scenarios where labeled data are extremely scarce. Under such conditions, gradients computed from the limited labeled examples may not adequately represent the overall data distribution, whereas gradients from unlabeled data could still convey useful information. Consequently, filtering the gradient directions from unlabeled data might prevent the model from learning from potentially informative signals and could exacerbate overfitting. It would therefore be valuable to include experimental validation under extremely low-label settings, for instance, with only one or two labeled examples per class, to better assess the robustness of the method in data-scarce environments.
3.Unclear explanation of some important details. It’s unclear how the empirical verification is conducted and how Eq. (2) is obtained.
4.Lack of more labeled ratios. Currently, the method is evaluated on only two labeled ratios for one dataset, while other datasets are tested under a single labeling ratio. To more comprehensively assess the robustness and practical applicability of the proposed approach, it would be beneficial to include experiments across a wider range of labeled data ratios.
References: [R1] Yang, Lihe, Zhen Zhao, and Hengshuang Zhao. “Unimatch v2: Pushing the limit of semi-supervised semantic segmentation.” IEEE Transactions on Pattern Analysis and Machine Intelligence 47.4 (2025): 3031-3048.
- 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
In Table 1, the ASD result of 2.19 achieved by BCP-SGD_CO should also be highlighted in bold since it obtains the same value as the proposed method.
- 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?
This paper makes a novel methodological contribution to semi-supervised segmentation by introducing a scheme to mitigate gradient noise from unlabeled data. Its effectiveness is convincingly demonstrated through extensive experiments across multiple classic semi-supervised methods and optimizers.
- 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.
Although concerns about diversity and theoretical security are not well addressed, this work introduces new insights into the semi-supervised community.
Author Feedback
We thank the reviewers for the constructive feedback. We address the main concerns below.
1.Experimental scope and SOTA positioning (R1.3, R2.4, R3.1). We evaluate an optimizer-level effect with a paired design — same pipeline, backbone, split, seed, and schedule; the optimizer is the only variable. We run this across two representative 3D SSL paradigms: BCP and CMT. In response to R1.3, we clarify that HCO was not evaluated only with BCP. For broader context, HCO is competitive with recent 3D medical SSL methods despite changing only the optimizer: on Pancreas-CT, it reaches 83.71 Dice, outperforming BCSI (AAAI’26, 83.24) and matching DuCiSC (TMI’25, 83.71); on LA, it reaches 90.75 Dice, comparable to AD-MT (ECCV’24, 90.55), PMT (ECCV’24, 90.81), and below BCSI (AAAI’26, 91.07). HCO is orthogonal to such architectural advances. UniMatch V2 (R3.1) is not an established 3D medical SSL protocol; we adhere to standard 3D medical setups and will state this explicitly in the revision.
2.Method justification and theory (R1.1–2, R2.1–2, R3.3).
Why binary alignment gating? HCO is designed for the asymmetric reliability structure of SSL: labeled gradients come from ground truth, whereas unlabeled gradients come from noisier pseudo-label/consistency signals. The goal is not to balance two equally trusted objectives but to prevent unlabeled signal from reversing the trusted labeled direction. Binary gating is the minimal mechanism that enforces this. Note: only the entry into momentum is binary; the resulting momentum and updates remain continuous-valued, preserving update diversity (R1.2). PCGrad/CAGrad handle gradient conflict symmetrically, while CO applies symmetric alignment-based masking. HCO differs from both by making the labeled signal a privileged reference. We ablate this source-aware hierarchy directly: across our result tables, replacing HCO with CO removes or substantially reduces the performance gains. We will sharpen this in the revision.
Theory. Eq. (2) is not a pointwise requirement at every step or coordinate. It follows by substituting HCO’s gated update direction into the standard Adam-style descent inequality; the only HCO-specific assumption is positive alignment of that gated direction in expectation. Fig. 1D supports this assumption by showing positive labeled–unlabeled alignment on average, while the negative tail motivates HCO’s filtering of conflicting unlabeled components. We will revise the text to clarify this distinction and the empirical verification procedure.
3.Fetal gains, low-label, and protocol robustness (R2.3, R2.5, R3.2, R3.4). R2.3 raises concern that fetal gains may reflect a setup or validation issue. Several design aspects argue against this:
- Paired design. Baseline and HCO use the same code, data, splits, backbone, and schedule; only the optimizer differs, reducing the risk that the gains come from pipeline differences.
- Headroom. The supervised-only baseline reaches only 51% Dice (lungs), 61% (liver) — far from ceiling, helping explain why an effective optimizer change yields larger absolute gains here than on more saturated benchmarks.
- 16 independent CMT runs. We train each model from scratch across 4 labeled × 4 unlabeled splits.
- No HCO tuning. We use CMT default hyperparameters; HCO introduces none.
This also addresses R3.2’s low-label concern: scarcity should be interpreted relative to task difficulty, not raw label count. With only five labeled fetal volumes, supervised performance is low (51%/61% Dice), yet HCO gives its largest gains (+24% lungs, +16% liver) — suggesting labeled-driven momentum remains a useful anchor even under severe effective scarcity. More broadly, we follow the published label ratios and protocols of BCP and CMT (R3.4): LA — BCP 4/76 and 8/72 splits with two seeds; Pancreas — BCP 12/50 split; Fetal — the 16 CMT folds above with CMT defaults. We will clarify these details in the revision (R2.5).
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 is borderline (scores 3–5): reviewers find the idea simple, practical, and promising, but question whether it is sufficiently justified and validated. The main issue is method grounding: the rebuttal should better explain why alignment-based gating is the right design (vs. alternatives) and strengthen the theoretical justification (assumptions, convergence). The other key concern is experimental support, including unusually large gains, limited comparisons (missing recent SOTA and related gradient-conflict methods), and insufficient validation (more seeds, clearer protocols, broader label ratios and settings, especially low-label regimes). Clarifying some implementation details would also help. Overall, most concerns seem addressable, and a focused rebuttal could strengthen the paper’s case for acceptance.
- 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 the rebuttal, all reviewers agreed that the paper should be accepted. The Area Chair supports this consensus and recommends acceptance.
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 carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper proposes Hierarchical Cautious Optimization, a simple and practical optimizer-level method for semi-supervised medical image segmentation that explicitly distinguishes the reliability of labeled and unlabeled gradients. The reviewers generally recognized the method as intuitive, easy to implement, and potentially useful as a plug-and-play component for existing momentum-based optimizers. Although the initial reviews raised concerns regarding theoretical depth, reliance on gradient-direction alignment, limited validation across stronger semi-supervised frameworks, and robustness under extremely low-label settings, the rebuttal was considered acceptable by the reviewers. Importantly, all submitted post-rebuttal opinions support acceptance, and the work offers an interesting perspective on mitigating noisy unlabeled gradients in semi-supervised learning. Therefore, I recommend acceptance, with the expectation that the authors will further clarify the theoretical assumptions, empirical verification details, and limitations of the alignment-based filtering strategy 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.
All reviewers have reached a consensus in favor of accepting the paper.
