Abstract

Multi-rater medical image segmentation reflects the inherent ambiguity of clinical interpretation, where experts often agree on the semantic extent of a target but differ in fine-grained boundary precision and textural cues. Existing approaches frequently collapse this variability into consensus labels or treat inter-rater differences as noise, leading to overconfident predictions and limited control over personalized outputs. In this work, we view rater variability as structured and preference-driven, modeling it through a style factor that captures individual expert delineation of ambiguous regions while preserving a shared content representation for anatomical consistency. We introduce a unified probabilistic framework that disentangles shared content from rater-specific style and employs lightweight rater conditioning to generate personalized yet anatomically consistent segmentations without decoder duplication. To better align learning with actual expert disagreement, we propose a novel disagreement-aware boundary loss that focuses supervision within a controllable band around regions of annotation divergence. This loss heightens sensitivity to contour-level mismatches where inter-rater ambiguity is greatest, while remaining stable in regions of agreement, thereby tying uncertainty estimation and personalization to clinically meaningful boundary disagreement. Extensive experiments on LIDC-IDRI and NPC-170 show that our method surpasses strong multi-rater baselines in both aggregated and per-rater performance, while producing diverse, plausible segmentations.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5829_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/sanazkarimi/style-aware

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KarSan_Probabilistic_MICCAI2026,
        author = { Karimijafarbigloo, Sanaz AND Kazerouni, Amirhossein AND Kheirkhah, Alireza AND Protserov, Sergey AND Azad, Reza AND Brudno, Michael AND Taati, Babak AND Samek, Wojciech AND Merhof, Dorit},
        title = { { Probabilistic Multi-rater Segmentation via Style-Aware Boundary Conditioning } },
        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

    This paper addresses multi-rater medical image segmentation by modeling inter-rater variability as structured style preferences rather than noise. It proposes a probabilistic framework with two core components: 1.Style-Conditioned Personalization (SCP), a lightweight module that performs fine-grained, rater-specific feature modulation without decoder duplication. 2.Disagreement-Aware Boundary Loss (DABL), which focuses supervision on ambiguous boundary regions where experts diverge. Experiments on LIDC-IDRI, NPC-170, and Kvasir-SEG show improved performance in rater-specific segmentation, distribution alignment, and uncertainty calibration over recent baselines.

  • 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.Clear clinical motivation: Modeling inter-rater variability as structured style is clinically meaningful and better than treating disagreement as noise. 2.Clear problem-solving design: SCP avoids freezing shared features and enables boundary-aware personalization, which is more suitable for fine-grained medical segmentation than existing methods like D-Persona. 3.Novel and effective loss design: DABL explicitly emphasizes regions of inter-rater disagreement, improving alignment between model uncertainty and clinical ambiguity. 4.Lightweight and efficient: SCP introduces only ~0.11M parameters. 5.Comprehensive and rigorous experiments: Validations cover multi-rater segmentation, distribution fitting, diversity, calibration, robustness to label noise, and ablations.

  • 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 causal validation for uncertainty alignment: Although uncertainty correlates with expert disagreement, the paper does not rigorously verify that this alignment is caused by SCP and DABL rather than probabilistic modeling or longer training. 2.Two-stage training is suboptimal: Freezing the probabilistic backbone in Phase II prevents end-to-end optimization and may limit representation alignment. No comparison with joint training is provided. 3.The paper lacks discussion on: cross-modality / cross-organ generalization, performance on extremely ambiguous or small lesions, and comparison with inter-rater human agreement (e. g. , Kappa).

  • 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?

    The paper addresses multi-rater medical image segmentation with a principled framework and comprehensive experiments. Although there remain some weaknesses in validation and analysis, the paper is acceptable for publication.

  • 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 paper addresses the problem of multi-rater variability in medical image segmentation from a different perspective than most prior work. Instead of treating inter-rater differences as noise or collapsing them into a consensus label, the authors model these differences as structured and meaningful, particularly along object boundaries.

    The main technical contribution is a style-conditioned personalization mechanism that injects rater-specific information directly into intermediate feature representations. This is a notable design choice, as many existing approaches either rely on shared latent spaces or apply personalization only at the output level. By conditioning features through learned rater embeddings (via gating and FiLM-like modulation), the model is able to produce segmentation outputs that reflect individual annotator preferences without duplicating the decoder. This leads to more flexible and fine-grained control over boundary behavior.

    A second contribution is the disagreement-aware boundary loss . The authors explicitly identify regions of inter-rater disagreement and place greater training emphasis on these areas. This is a reasonable and well-motivated design, as standard losses tend to overweight large homogeneous regions and under-emphasize clinically ambiguous boundaries. By focusing supervision on disagreement regions, the method better aligns optimization with the actual source of annotation variability.

    These components are integrated within a probabilistic segmentation framework based on latent variable modeling, allowing the network to generate diverse and plausible outputs rather than a single deterministic prediction. Empirically, the method shows improvements in both distribution-level metrics (e.g., GED) and rater-specific performance, suggesting that it captures both uncertainty and personalization more effectively than prior approaches.

    Overall, the key contribution is a shift in how multi-rater variability is handled: from being treated as noise to being explicitly modeled as a structured, learnable factor that can be controlled and reproduced at inference time.

  • 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.

    A key strength of this work is its conceptual shift in how multi-rater annotations are modeled. Instead of collapsing annotations into a consensus or treating disagreement as label noise, the paper explicitly models variability as a structured, preference-driven phenomenon. This is operationalized through a disentangled representation of shared anatomical content and rater-specific “style,” which is both intuitive and clinically meaningful. This formulation aligns well with how experts actually differ in practice, primarily along boundary interpretation, making the approach more promising to real-world annotation processes

    The proposed Style-Conditioned Personalization is a notable technical contribution. Unlike prior approaches that rely on separate heads or post-hoc adaptation, SCP conditions intermediate feature representations using rater-specific embeddings. This enables fine-grained, rater-dependent adjustments during decoding, allowing the model to recover subtle boundary variations that would otherwise be lost in a shared representation. Importantly, this is achieved without duplicating the decoder, making the design both expressive and efficient .

    Another key strength is the Disagreement-Aware Boundary Loss , which focuses supervision on regions where annotators disagree. By constructing a boundary band around disagreement areas, the method emphasizes clinically meaningful uncertainty while maintaining stability in regions of agreement. This aligns the learning objective with the true source of ambiguity in medical segmentation tasks .

    The integration of probabilistic modeling with personalization is also well executed. The latent space captures distributional ambiguity, while the style mechanism encodes annotator-specific preferences, resulting in diverse yet controlled predictions. Empirically, the method demonstrates consistent improvements across multiple datasets and metrics, including both diversity and per-rater accuracy

    Finally, the approach remains lightweight, adding minimal computational overhead, and shows strong uncertainty calibration aligned with inter-rater disagreement, which is particularly important for clinical applicability.

  • 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 provides limited evidence of clinical impact. Experiments are conducted on standard benchmarks without demonstrating how personalized outputs would influence real clinical workflows. Moreover, the reported failure cases suggest a bias toward majority annotations, potentially limiting the model’s ability to capture minority but clinically relevant interpretations. Also, novelty is another concern, since the method mainly builds on existing probabilistic U-Net and multi-rater personalization frameworks, with the main contributions coming from extensions in conditioning and loss design rather than a fundamentally new approach.

    Some of the detailed comments here:

    In Eq. (2), the rater-specific modulation is defined as a Gaussian-like excitation. There is no justification that such a distribution adequately captures complex annotator behaviors, which may be multi-modal or spatially varying.

    For DABL steps, these steps introduce hyperparameters (τ,λ,α) that directly influence optimization but are not systematically analyzed

    The two-stage optimization in Eq. (7) where the backbone is frozen before training SCP. This may prevent gradients from DABL and personalization from refining earlier feature representations, resulting in suboptimal alignment between shared and rater-specific components

  • 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?

    I think this is a solid paper overall. The idea of modeling rater-specific differences, especially around boundaries, makes sense and is relevant for medical segmentation. The SCP module and the disagreement-aware loss are both reasonable extensions, and the method is built on a sound probabilistic framework.

    The experiments are fairly convincing. The paper shows consistent improvements over strong baselines across multiple datasets, and the evaluation is thorough with different metrics and ablations. It’s clear the method works and is not just a small tweak.

    That said, I don’t think the novelty is very high. A lot of the work builds on existing probabilistic U-Net and personalization ideas, and the contributions feel more like incremental improvements rather than a fundamentally new direction. Also, some parts (like the style tokens and modulation) could be explained more clearly as right now it takes some effort to fully understand the intuition.

    Another issue is that there’s no code or discussion of reproducibility, and comparisons with some more recent methods could be stronger.

    Overall, it’s a good piece of work with solid results, but not quite strong enough for a clear accept, so I’d place it slightly above the acceptance threshold.

  • 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 a novel probabilistic framework for multi-rater medical image segmentation that explicitly models differences between expert annotators as meaningful variation rather than noise. Its main contribution is the introduction of style-conditioned personalization (SCP), which disentangles shared anatomical content from rater-specific boundary “style” and conditions intermediate features to generate personalized segmentations without duplicating the model, along with a disagreement-aware boundary loss (DABL) that focuses learning on regions where annotators disagree (typically boundaries). Built on a Probabilistic U-Net backbone, this approach produces diverse, anatomically consistent, and rater-specific outputs while improving uncertainty calibration and segmentation accuracy across multiple datasets.

  • 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’s main strengths lie in its principled treatment of inter-rater variability as structured information rather than noise, enabled by a well-designed style-conditioned personalization (SCP) module that captures fine-grained, rater-specific boundary preferences without requiring heavy architectural duplication, making it both effective and parameter-efficient; its novel disagreement-aware boundary loss (DABL), which focuses supervision on clinically meaningful regions of annotator disagreement and improves boundary precision and uncertainty alignment; and its integration within a probabilistic framework that generates diverse yet anatomically consistent segmentations with well-calibrated uncertainty. Additionally, the method demonstrates consistent empirical gains over strong baselines across multiple datasets, while maintaining minimal computational overhead and producing outputs that better reflect real expert variability.

  • 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.

    While the proposed SCP module improves personalization, the method still relies on a two-stage training procedure with partial freezing, which may limit full end-to-end optimization and adaptation of shared features to rater-specific nuances; the approach introduces additional design complexity (e.g., style tokens, FiLM modulation, boundary weighting) without fully isolating which components contribute most, and ablations remain somewhat limited; the evaluation, although strong, is confined to a few benchmarks and lacks deeper analysis of generalization to unseen raters or domains, which is critical for real clinical deployment; the method assumes access to multiple annotations per image during training, restricting applicability in annotation-scarce settings; and finally, while DABL focuses on disagreement regions, it may overemphasize boundaries at the expense of interior consistency, and its hyperparameters (e.g., band width, weighting) require tuning and are not extensively studied.

  • 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?

    See comments above

  • 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

N/A




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.

    All reviewers recommended acceptance of this work, recognizing its clear motivation, reasonable technical design, and superior experimental performance. The authors are encouraged to address the reviewers’ comments in the camera-ready version.



back to top