List of Papers Browse by Subject Areas Author List
Abstract
Reliable pixel-level uncertainty quantification holds the potential to transform clinical workflows by enabling high-fidelity longitudinal monitoring and distinguishing true pathological changes from artifacts. Ideally, these models provide the stability required for critical treatment planning and surgical intervention. However, standard deep learning models often suffer from miscalibration, yielding overconfident predictions that mask underlying vulnerabilities at subtle pathological boundaries. To address this, we propose QUAM-SM, a post-hoc framework using targeted adversarial search to identify “adversarially fragile” pixels. By actively seeking perturbations that expose predictive instability, our method highlights regions where decisions are most vulnerable to being flipped. Importantly, the framework disentangles epistemic uncertainty from aleatoric uncertainty. Experiments on two public datasets with multiple expert annotations demonstrate that QUAM-SM outperforms both standard and recent uncertainty estimation approaches in terms of reliability and boundary sensitivity. Code is available at https://github.com/HanaJebril/quam_sm
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3376_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HanaJebril/quam_sm
Link to the Dataset(s)
REFUGE dataset: https://refuge.grand-challenge.org/
QUBIQ2021 dataset: https://qubiq21.grand-challenge.org/
BibTex
@InProceedings{JebHan_Quantification_MICCAI2026,
author = { Jebril, Hana AND Pinetz, Thomas AND Klambauer, Günter AND Bogunović, Hrvoje},
title = { { Quantification of Uncertainty with Adversarial Models in Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes QUAM-SM, a post-hoc adversarial framework for pixel-level uncertainty quantification in medical image segmentation. The method identifies “adversarially fragile” pixels by performing targeted adversarial search around a fixed pre-trained model and quantifies uncertainty based on prediction instability. The authors further claim disentanglement of epistemic and aleatoric uncertainty and demonstrate improved alignment with inter-rater variability on multi-annotator 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.
1.The problem is highly relevant to medical image analysis, particularly for safety-critical applications where uncertainty estimation is essential. 2.The idea of leveraging adversarial perturbations to probe predictive instability is interesting and intuitively aligned with identifying decision boundary sensitivity. T 3.he use of multi-observer datasets and the attempt to relate uncertainty to inter-rater variability are also commendable and clinically meaningful.
- 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 methodological novelty is limited. The core idea of using adversarial perturbations to estimate uncertainty has been explored in prior work, and the proposed formulation appears to be an incremental extension without sufficient theoretical grounding. 2.The difference between epistemic and aleatoric uncertainty is claimed but not rigorously justified. The reliance on access to training data for the penalty term weakens the “post-hoc” claim and limits applicability in realistic deployment settings. 3.The adversarial search formulation lacks clarity and reprodcibility: the use of multiple targeted attacks, their design choices and mixed importance sampling is mentioned but not clearly integrated into a coherent framework. Computational cost is also a concern, yet no runtime analysis is provided. 4.Comparisons with strong uncertainty baselines (e. g. , deep ensembles, test-time augmentation, diffusion-based methods) are either missing or insufficient. The evaluation of uncertainty quality (e. g. , calibration metrics, risk-coverage analysis) is limited, and more rigorous analysis would strengthen the claims. Some important comparisons are missing, for instance (a few):
- Kohl et al. , “A probabilistic U-Net for segmentation of ambiguous images,” NeurIPS 2018
- Baumgartner et al. , “Visual feature attribution using Wasserstein GANs,” CVPR (uncertainty near boundaries) Ho et al. , “Denoising Diffusion Probabilistic Models,” NeurIPS 2020 Sensoy et al. (Evidential DL) 5.Statistical evaluations are missing. 6.The effect of iterations shows that the performance improves as the number of iterations increase. However, it would be good to know the upper cap providing information on when to stop.
- 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 provide sufficient information for 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 work would benefit from stronger theoretical justification, comprehensive comparisons, and analysis of computational feasibility before being considered for acceptance. Several strong and widely used uncertainty estimation approaches are missing, including test-time augmentation methods (Wang et al., MICCAI 2019), evidential/Dirichlet-based models (Sensoy et al., NeurIPS 2018), and recent diffusion-based uncertainty frameworks.
- 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.
My points have been addressed almost except for statistical evaluation. So I am inclined towards accept.
Review #2
- Please describe the contribution of the paper
This paper proposes QUAM-SM, a post-hoc uncertainty quantification framework for medical image segmentation based on adversarial model search. The method extends published QUAM for segmentation to identify “adversarially fragile” pixels whose predictions can be easily flipped under targeted adversarial perturbations in model space. Experiments on REFUGE and QUBIQ2021 show improved correlation with multi-annotator entropy maps compared with MC Dropout, Deep Ensembles, Probabilistic U-Net, and SURE.
- 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 addresses an important problem: reliable pixel-wise uncertainty estimation for medical image segmentation. 2.The extension of QUAM using adversarial fragility for uncertainty is interesting in the segmentation setting. 3.The morphology-based target masks are well aligned with segmentation boundary ambiguity and appear empirically effective. 4.The paper includes ablation studies on target mask design, kernel size, and search iterations.
- 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.
This work appears to extend QUAM from classification to segmentation. In this context, the presentation of the method could be improved by focusing more explicitly on the segmentation-specific modifications relative to the original QUAM framework. At present, a substantial portion of the method description seems similar to the original formulation, while the task-specific novelties are not sufficiently emphasized. A more concise presentation of the inherited components, together with a clearer articulation of the new technical elements introduced for segmentation, would help clarify the contribution.
Given the critical nature of the target application, the paper would benefit from a more systematic evaluation of uncertainty quality. In particular, the use of multi-annotator entropy maps as a validation target requires stronger justification. Such entropy maps appear to primarily reflect annotator disagreement and may therefore be more closely associated with aleatoric uncertainty, while also depending on factors such as the annotation protocol, the number of annotators, and potential annotator bias, etc. A more careful discussion of these limitations, as well as of what aspect of uncertainty this surrogate target is intended to capture, would strengthen the experimental section. Alternatively, the authors may consider an evaluation strategy more closely aligned with the original QUAM work, for example by assessing uncertainty quality through OOD sample detection.
The paper would also benefit from a more detailed description of the dropout-based ensemble baselines. While such methods may underestimate uncertainty due to overconfident predictions, their behavior can depend substantially on the experimental setup and hyperparameter choices. Since the empirical evaluation is central to the paper, a clearer justification of the baseline configuration would help ensure that the comparison is fair and convincing.
- 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 provide sufficient information for 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?
Overall, I find the paper to be borderline in terms of contribution. The idea of extending QUAM to segmentation is interesting and meaningful, particularly given the importance of the target application. However, a direct extension to a new task does not by itself establish sufficient technical merit unless the novel methodological contributions are clearly articulated. Moreover, for an application-focused paper, the current experimental validation is not rigorously designed and not yet sufficient to convincingly demonstrate the quality of the estimates in terms of either the aleatoric or epidemic uncertainty or to clarify how these estimates can be used in practice.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper proposes QUAM-SM, a post hoc uncertainty quantification framework for medical image segmentation that identifies adversarially fragile pixels by optimizing alternative model weights on a per-test-image basis. The method starts from a fixed reference segmentation, then performs targeted adversarial model search to generate alternative plausible predictions that diverge from the reference while remaining constrained by a penalty tied to training performance. The resulting set of adversarially found models is aggregated using a weighted Bayesian-model-averaging-style scheme, and total, aleatoric, and epistemic uncertainty are computed from the weighted predictions. The method is evaluated on multi-annotator datasets and is claimed to better align predicted uncertainty with inter-rater entropy.
- 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) Rather than relying only on stochastic posterior approximations such as MC Dropout or ensembles, the paper explicitly searches for vulnerable predictions through adversarial model optimization. This is a creative and potentially impactful angle on uncertainty estimation. 2) The problem is clinically relevant and clearly within the core MICCAI scope: pixel-level uncertainty in medical image segmentation, with motivation tied to boundary sensitivity, treatment planning, and multi-observer ambiguity. 3) Using REFUGE and QUBIQ2021, both with multiple expert annotations, is a meaningful choice because multi-observer entropy provides a plausible reference for uncertainty analysis, especially for aleatoric uncertainty. 4) Table 1 shows substantial improvements over several well-known baselines across epistemic, aleatoric, and total uncertainty, especially on REFUGE. The qualitative examples on page 7 also visually support the claim that the method is more boundary-sensitive. 5) The paper compares different adversarial target strategies and morphology settings, and studies the effect of the number of iterations M. These ablations give some insight into what aspects of the design matter. 6) The paper clearly distinguishes total, aleatoric, and epistemic components using a weighted ensemble-style decomposition. This makes the contribution more substantive than just producing a single uncertainty heatmap.
- 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 paper refers to the weighting scheme as Mixture Importance Sampling (MIS) and claims a “statistically consistent estimate,” but the adversarially optimized model weights are not obviously sampled from a known proposal distribution in the standard importance-sampling sense. As written, the weights look more like heuristic softmax reweighting based on penalty loss than a rigorously derived MIS estimator. This is a central technical concern. 2) The penalty loss depends on training data during post-hoc adversarial search. This is acknowledged in the paper, but it is a substantial deployment limitation for a method presented as a post hoc uncertainty estimator. In realistic clinical settings, retaining and repeatedly using the full labeled training set at inference time may be impractical or undesirable. 3) Important algorithmic details are missing or under-specified. Several elements are not described clearly enough for reproduction or full technical assessment. 4) Fairness of baseline comparison is somewhat unclear. The proposed method uses weighted aggregation of adversarially found models, while the baselines are presented “without weighting.” 5) The paper computes separate aleatoric and epistemic maps, but most of the validation still relies on inter-rater entropy, which is primarily an aleatoric proxy. 6) Since the method optimizes adversarial weights per test image and accesses training data at inference time, computational cost is likely a major practical issue. The lack of runtime reporting is a notable omission for a potentially expensive method. 7) There are small inconsistencies, such as QUBIQ/QUBIC naming, overloaded notation, and ambiguity around why SDice appears in uncertainty-specific columns in Table 1.
- 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 provide sufficient information for 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
This paper explores a novel and interesting direction for pixel-wise uncertainty estimation in segmentation. I appreciate the attempt to move beyond passive stochastic sampling and instead actively probe model fragility through adversarial search. The use of multi-annotator datasets is also a meaningful strength. Overall, I find the idea promising and fairly original, but the paper currently needs more methodological rigor and clearer practical validation.
- 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 am not convinced that the current submission provides enough rigor for acceptance. The theoretical framing of the weighting scheme as mixture importance sampling appears too strong relative to the description provided; several core algorithmic details are missing, the method requires access to labeled training data at test time, and the evaluation focuses mainly on correlation with inter-rater entropy rather than more direct uncertainty-utility metrics. These issues do not make the paper uninteresting, but they do leave too many open questions at present.
- Reviewer confidence
Very confident (4)
- [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
Paper ID 3376 We thank the Meta-Reviewer and Reviewers for their constructive feedback. We appreciate the reviewers’ recognition of the clinical relevance of our work, particularly for safety-critical applications.
Stronger Experimental Support(MR, R1, R2) While we agree with the reviewers that stronger experimental support is always preferable, we specifically used datasets with multiple annotations. However they are too small for diffusion-based approaches, which would require a larger external training dataset making the comparison unfair. We would like to stress that the original submission does include comparisons to Probabilistic U-Net ([8] Kohl et al) and a recent evidential/dirichlet-based approach (SURE [11]), while the test-time augmentation (TTA) results are as follows: Results (REFUGE): QUAM-SM achieves a Total PCC of 0.79, significantly outperforming TTA (0.59) Results (QUBIQ): QUAM-SM maintains a Total PCC of 0.58, outperforming TTA (0.57). We used grid search for the baseline hyperparameters and will clarify this in the paper. While the weighting does improve the results, even our unweighted method improves upon the baselines.
Runtime(MR, R1, R3) While the post-hoc adversarial search requires 260s of test-time compute compared to the 2s inference of stochastic methods, this is still doable within routine clinical practice, and we argue that in safety-critical medical applications, diagnostic reliability is more vital than near-instantaneous latency. Efficiency and Minimal Data Dependency(MR, R1, R3): While our ablation study on the number of iterations (M), Table 2 part 3, demonstrates that the performance is correlated with the number of training batches used in the penalty term, we can also outperform state-of-the-art with a small number of images (M=2 iterations, utilizing only 2% of the training data). Still, some subset of training data is required, which serves as a functional advantage to increase trustworthiness. Though regulatory-approved software does require keeping the training set on record.
Mixture Importance Sampling (MIS) Rigor(MR, R3) We clarify that QUAM-SM utilizes MIS, following the theoretical foundation of the original QUAM framework. In particular, the weight matrices W_i are exactly the same as in the original QUAM work and therefore, the theoretical results that they claim, such as statistical consistency, hold for our setup as well. We will rewrite this section in the camera-ready version to improve clarity.
Validation Target(MR,R2,R3) We use the most widely used decomposition of uncertainty into aleatoric and epistemic as derived and justified in [Depeweg2018; Gal2017; Hüllermeyer2021]. We will include those references. We utilize multi-annotator entropy based on the clinical conjecture that expert disagreement is the most “valid” uncertainty in a diagnostic setting. Our results demonstrate that this ambiguity is primarily captured by the aleatoric component, as is expected. Furthermore, we showcase that our approach consistently works best for this important clinical use-case. Our provided maps align with the inherent ambiguity recognized by human experts, which is required for safe treatment planning. We will clarify this in the text.
Technical Novelty(MR,R2) The original QUAM work only considers classification problems, where adversarial attacks are easy to define, as it is essentially a label flip. For dense labels such as segmentation masks, these adversarial attacks could in theory be single pixel flips or global pixel flips and it is a priori not clear what works well. We extend QUAM by aligning the chosen pixels with boundary ambiguity as noted by R2.This design choice leads to significant improvements over the naive “Global Extreme Masks” by +14%. As far as we know, this is the first application of adversarial perturbation to segmentation.
Reproducibility(MR,R3) To support full reproducibility, we prepared a code repository that will be uploaded upon acceptance.
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 reviewers raise substantial concerns about limited novelty and positioning, as the work appears to extend QUAM from classification to segmentation while relying on previously explored ideas such as adversarial perturbation-based uncertainty, without clearly articulating the segmentation-specific technical contribution. The authors should address major methodological concerns regarding the claimed separation of epistemic and aleatoric uncertainty, the use of training data during “post-hoc” inference, the validity of the Mixture Importance Sampling formulation, and missing details needed for reproducibility. The rebuttal should also provide stronger experimental support, including comparisons with stronger uncertainty baselines, calibration/risk-coverage or OOD-style uncertainty evaluation, statistical analysis, runtime reporting, and clearer justification of multi-annotator entropy as an uncertainty validation target.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper is accepted as a borderline but worthwhile contribution. The work is creative and technically interesting, particularly in extending QUAM-style adversarial uncertainty estimation to medical image segmentation, and the rebuttal addressed most major concerns, including comparisons and algorithmic clarification. Although some limitations remain regarding statistical evaluation, theoretical framing, and the need for labeled training data at test time, the paper has sufficient novelty and potential to stimulate discussion on uncertainty estimation for segmentation.
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.
Only one reviewer submitted a post-rebuttal recommendation, which changed an initial weak reject to an accept, mentioning that almost all the raised concerns were addressed in the rebuttal. That leaves two reviewers, one leaning to accept and one leaning to reject in the original review, where the latter demonstrated a lower confidence (2). Having looked at the paper, the reviews and the rebuttal, it seems that this paper has sufficient merits to warrant an accept.
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.
The reviewers appreciate the significance of the problem and extensive experimental results, and the initial concerns have been mostly addressed during the rebuttal. I recommend accepting this submission.