List of Papers Browse by Subject Areas Author List
Abstract
Learning robust medical image segmentation models is challenging when training datasets contain mixed-quality annotations arising from inter-observer variability and ambiguous boundaries. Although existing methods address noisy supervision through label correction or supervision selection strategies, they rarely integrate both, which limits effective noise handling. In response, we propose RGSS, Reliability-Guided Supervision for medical image Segmentation under annotations of varying qualities. RGSS jointly estimates annotation reliability via co-training and Gaussian Mixture Model (GMM), and adaptively reconstructs supervision targets using a pretrained diffusion-based mask refiner. The refinement strength is adjusted by the estimated noise severity, enabling aggressive correction for heavily corrupted masks while preserving structural details in mildly noisy cases. To mitigate confirmation bias and prior-induced errors, refined masks are further blended with peer-model predictions through epoch-dependent scheduling. Extensive experiments on two public medical segmentation benchmarks with four backbone networks demonstrate that our method consistently improves robustness and segmentation accuracy under varying levels of annotation noise, outperforming existing baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2250_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)
Kvasir-SEG dataset: https://datasets.simula.no/kvasir-seg/
ISIC 2017 dataset: https://challenge.isic-archive.com/data/#2017
BibTex
@InProceedings{JeoMin_Learning_MICCAI2026,
author = { Jeong, Minjae AND Lee, Seungjoo AND Lee, Jaejin AND Kim, Won Hwa},
title = { { Learning Robust Medical Image Segmentation Under Mixed-Quality Annotations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a co-training approach to medical image segmentation by deliberately including noise during training to demonstrate its effect on unreliable annotations, a common issue in the field. They compare this against other segmentation styles, showing a slight but consistent improvement.
- 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 strengths, The paper breaks down the method, explains the key sections, and shows how to reproduce the technique step-by-step. If you allow the code to be released as well, the paper would be fully reproducible. The paper has a clear progression upon existing techniques, where you have built and improved in a logical way. The paper flows well, guiding the reader through each stage of your work. Some parts are a bit high-level for a standard reader, but overall it is well-executed. The comparison against similar techniques on a wide range of noisy samples, and the metric usage makes sense. Fig 3 makes good use of showing under and over segmentations; expanding that onto Fig 4 would be an excellent showcase of the segmentation changes.
- 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.
Paper weakness, The paper formatting is very odd and does not look like a MICCAI paper. The break down into the subsection are clear, but it doesn’t meet the MICCAI standard. Likewise, tables are not in the MICCAI standard. There are some up-to-date references, but some more would also help, as many are older.
- 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
Overall, You have done some really good work; the technique and results make sense and are logical, and the reproducibility is great. Other than the oddness of the formatting, not the flow, the paper would be a good addition.
- 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?
The paper, overall, is a good contribution. The method can assist others in making more reliable segmentation under an array of noisy annotations, which is a common issue in MIA.
- 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
This paper proposes a novel iterative framework, RGSS, for robust medical image segmentation under mixed-quality annotations. It estimates annotation reliability through co-training and Gaussian mixture models, adaptively refines noise masks using a reliability-guided diffusion refiner, constructs robust supervision by combining the refined mask with peer model predictions to mitigate confirmation bias, and achieves excellent segmentation accuracy and generalization ability on two public benchmarks and four mainstream backbone networks.
- 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 paper proposes a novel RGSS framework that unifies label reliability estimation and adaptive masking refinement, addressing the key limitations of existing methods that rely solely on label correction or supervised selection. 2.This paper designs an innovative reliability-guided adaptive diffusion refinement method. This refinement method dynamically adjusts the correction intensity based on the sample noise severity estimated by co-training and GMM, rather than using a fixed-intensity refinement, thus better preserving model structure and improving accuracy. 3.This paper introduces peer-guided supervised fusion and round-related scheduling, effectively mitigating confirmation bias and prior error. This is a practical and effective design for robust learning under noisy labeling. 4.Comprehensive and rigorous evaluation: Validated on two publicly available medical segmentation benchmark datasets, four mainstream backbone networks, and three noise levels, the results show that the proposed method offers a sustained and significant performance improvement over robust baseline models. 5.The framework is model-agnostic and plug-and-play, easily integrated into existing segmentation workflows, and demonstrates high practical value for real-world clinical data with mixed-quality labeling.
- 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.Limited Experimental Validation: Only binary classification segmentation was tested, and the generalization ability of the method in multi-class medical image segmentation was not verified, which weakens its practicality. 2.Insufficient Ablation Experiments: The effectiveness of the core module was only verified under a single noise setting, but sensitivity analysis of important hyperparameters such as the reliability threshold τ, the initial number of training epochs e₀, and the diffusion step size range is lacking, making the stability of the method unclear. 3.Lack of Comparison with Recent State-of-the-Art Methods: This paper only compares with several traditional baseline methods, without comparing with the latest noise label segmentation methods proposed in the last two years. Therefore, the performance improvement of RGSS is not fully demonstrated. 4.Insufficient Model Efficiency Analysis: This paper does not report training time, computational complexity, or inference speed compared with baseline methods, which are important details for clinical applications and method reproducibility. 5.Incomplete Visualization and Analysis: The qualitative results only show a portion of the cases and do not provide a detailed comparison of the boundary refinement effect under different noise levels, thus failing to fully reflect the advantages of adaptive diffusion processing.
- 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.
(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 reliable RGSS method for medical image segmentation under mixed quality annotation, demonstrating significant innovation and solid experimental results. However, this method has some limitations in terms of experimental scope and detailed analysis. 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.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes RGSS (Reliability Guided Supervision for Segmentation), a method to handle mixed-quality annotations and noisy labels in medical image segmentation. The method explicitly models inter-observer variability and boundary ambiguity by utilizing a co-training framework with a Gaussian Mixture Model (GMM) to estimate annotation reliability. It then adaptively refines the noisy targets using a pre-trained diffusion-based mask refiner, followed by epoch-dependent blending with peer-model predictions to mitigate confirmation bias.
- 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 explicitly identifies and attempts to model a very real problem in medical image analysis: inter-observer variability and boundary ambiguity. The experimental evaluations are relatively comprehensive, covering multiple datasets and various backbone networks. The paper is clearly written, logically structured, and the proposed technical pipeline (Co-training + Adaptive Diffusion Refinement + Blending) is methodologically sound from a general machine learning perspective.
- 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.
Despite the solid methodological design, there are significant concerns regarding the realism and practicality of this method in the medical imaging domain:
Definition of Ground Truth (GT) in Clinical Practice: In real-world medical applications, if physician annotations are treated as inherently noisy and unreliable by the algorithm, what constitutes the actual ground truth? Without considering physician annotations (or a consensus of them) as the gold standard, it becomes philosophically and practically challenging to properly evaluate the effectiveness of the proposed method in a true clinical setting. Unrealistic Noise Simulation: In the experiments, annotation errors are primarily simulated using Gaussian noise and random morphological perturbations (e.g., random dilation and erosion). However, these synthetic settings differ significantly from real clinical annotation biases. They fail to capture the complex, structured differences arising from distinct disease topologies, ambiguous pathological borders, and varying levels of physician expertise. Overall, while the method is technically interesting, it appears more suitable for natural images or general machine learning tasks where “noise” is mathematically simpler. Its reliability and fundamental premise in the complex medical imaging domain remain questionable.
- 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
The technical pipeline is well-designed, but the gap between synthetic morphological noise and actual clinical inter-observer variability is large. If the authors could demonstrate the method’s effectiveness purely on datasets with multiple real human annotators (e.g., modeling the bias between a junior resident and a senior attending) rather than relying heavily on synthetically perturbed masks, the clinical value of this work would be much more convincing.
- 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?
My recommendation is a Weak Reject. The major factors for this decision are the concerns over the clinical realism of the problem formulation. While the paper is well-written and the experiments are extensive, the simulation of annotation errors using random morphological perturbations does not accurately reflect real-world medical annotation biases. Furthermore, the conceptual approach of treating physician annotations as “noisy” raises fundamental questions about how ground truth is defined and evaluated in medical imaging. The method feels more like a general computer vision technique applied to medical data rather than a solution tailored to real clinical complexities.
- 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
Author Feedback
We sincerely thank the reviewers for their constructive and encouraging comments. In our final manuscript, all major concerns raised in the reviews will be fully addressed. R1: Q) Formatting and Reference Issues A) Thank you for mentioning the need for additional up-to-date references. We will incorporate this feedback into our final version. In terms of formatting, our paper follows the standard MICCAI LATEX template. We will further confirm if any formatting discrepancies have occurred and adjust accordingly. R2-W1: Q) Extension to Multi-class Segmentation A) The proposed RGSS framework is extendable to multi-class segmentation. In particular, the reliability estimation and target reconstruction pipeline are agnostic to the number of classes, whereas the Bernoulli-based diffusion refiner can be replaced with multinomial-discrete diffusion formulations (e.g., D3PM[1]) for multi-class mask refinement. Additional multi-class validation will be included in the extended journal version. R2-W2: Q) Additional Experiments and Analysis Additional ablation studies across multiple hyperparameter settings, as well as efficiency and boundary-level analyses, were conducted during development but could not be included due to space limitations. We agree that these analyses are important for demonstrating the robustness, stability, and effectiveness of the proposed framework. These results and visualizations will be included in journal extensions of this work. R2-W3: Q) Insufficient Baselines A) We have selected representative baselines that are closely related to noisy-label medical segmentation and reproducible under our setting. To the best of our knowledge, there are currently limited recent methods with directly comparable settings for mixed-quality noisy medical segmentation. Nevertheless, we will further investigate and include additional recent methods and related works in the revised manuscript. R3-W1: Q) Annotation Validness in Clinical Practice A) Our framework does not disregard physician annotations as clinically meaningful supervision. Rather, we assume that medical annotations may exhibit varying levels of reliability due to inter-observer variability and boundary ambiguity, which are commonly observed in real clinical practice. Therefore, the objective of RGSS is not to replace physician supervision, but to alleviate inconsistencies among annotations by adaptively estimating annotation reliability and refining potentially corrupted regions. In our experiments, the original physician-provided annotations are still treated as the reference standard for evaluation, while our method aims to improve robustness against inevitable annotation uncertainty. R3-W2: Q) Clinical Validity of Noise Simulation A) Prior studies report that medical segmentation disagreement commonly manifests as boundary-level inconsistency and contour ambiguity [2,3]. Accordingly, our simulation applies morphology-based structural perturbations to approximate boundary inconsistency and local contour variation. Nevertheless, we agree that the proposed simulation does not fully capture all complex real-world clinical biases, such as systematic differences caused by physician expertise, disease-specific topology, or institution-level annotation protocols. We consider this an important direction for future work and will further investigate more clinically-grounded noise modeling strategies. [1] Austin, Jacob, et al. “Structured denoising diffusion models in discrete state-spaces.” NeurIPS (2021) [2] Zhang, Le, et al. “Disentangling human error from ground truth in segmentation of medical images.” NeurIPS (2020) [3] Han, Kai, et al. “Region uncertainty estimation for medical image segmentation with noisy labels.” IEEE Transactions on Medical Imaging (2025).
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Focusing on the critical problem of mixed-quality annotations, this work proposes a reasonable framework and achieves improved performance, as recognized by all reviewers. Considering its practical relevance and good contributions, this work is recommended for provisional acceptance. The authors are encouraged to address the reviewers’ comments in the camera-ready version.