List of Papers Browse by Subject Areas Author List
Abstract
Few-shot medical image segmentation methods typically assume a single ground-truth annotation, overlooking systematic variability across expert raters commonly observed in clinical datasets. We propose an attention-based prototype calibration framework for multi-rater few-shot segmentation that models rater-specific deviations from a consensus representation in prototype space. A lightweight yet principled attention operator directly refines rater prototypes without modifying the backbone feature extractor, making the approach fully compatible with existing prototype-based few-shot segmentation methods. This design preserves semantic consistency while enabling personalized segmentation outputs with minimal computational overhead. Experiments on multi-rater medical imaging datasets demonstrate consistent improvements over baseline prototype approaches, highlighting the effectiveness of structured prototype calibration for modeling annotation variability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3792_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/truong2710-cyber/JAPC
Link to the Dataset(s)
CURVAS dataset: https://zenodo.org/records/13767408
QUBIQ Brain-Growth dataset: https://qubiq21.grand-challenge.org/participation/
BibTex
@InProceedings{VuTru_AttentionBased_MICCAI2026,
author = { Vu, Truong AND Ho, Minh Khoi AND Xie, Yutong},
title = { { Attention-Based Prototype Calibration for Multi-Rater Few-Shot Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work formalizes few-shot multi-rater medical image segmentation and proposes a prototype-based personalization framework (JAPC) that models structured inter-rater deviations from a consensus prototype via joint attention. The framework incorporates pseudo-style generation to synthesize rater-style variants from style-agnostic pseudo labels, a calibration regularization loss to prevent prototype drift, and a two-stage training strategy for optimization stability. Experiments on CURVAS and QUBIQ demonstrate consistent improvements compared to 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) The methodology is well-motivated and sound. The JAPC module addresses the core limitation of existing few-shot methods by modeling structured inter-rater deviations from a consensus prototype via joint attention. (2) Evaluation on both CURVAS and QUBIQ demonstrates consistently strong efficacy across modalities, anatomical structures, and raters. (3) The problem formalization is neat and well-scoped, extending the standard episodic few-shot protocol to the multi-rater setting. (4) The ablation study is comprehensive and mirrors the motivation closely.
- 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 ablation study is comprehensive, but I would appreciate clarification on what the “negative control” is when a component is off. For instance, when “Attn” is ablated, it is ambiguous to me whether the prototype calibration module is replaced with identity pass-through (rater prototypes used as-is), replaced with standard self-attention (without the deviation-as-keys design), or replaced with simple averaging across rater prototypes. (2) Is this framework generalizable to 3D volumetric processing? It appears architecture-agnostic and could in principle be plugged into a few-shot supervoxel backbone. Some discussion or experiments with 3D backbones would strengthen the evaluation. (3) Both backbones (SSL-ALPNet, DSPNet) seem to share the same feature extractor. If architecturally distinct backbones (e. g. transformer ones) are available, they may be worth including. (4) If the official test set is released for QUBIQ Brain-Growth or if other options with larger test sets are available, they may be worth incorporating.
- 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?
I tentatively recommend acceptance. However, I would appreciate further clarification on the ablation settings and a larger test set for Brain MRI evaluation or a better characterization of statistical significance.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I am satisfied with the clarifications on ablation setup and the commitment to including further evidence of effective multi-rater modeling. Please ensure that the camera ready version clearly illustrates how the model captures annotator-specific styles, and the realism of the pseudo-style generation.
Review #2
- Please describe the contribution of the paper
In this work the authors propose a framework for few-shot multi-rater segmentation. Their main contributions are an attention-based prototype calibration module, a calibration loss to stabilize prototype regularization, a pseudo-style generation scheme to model multi-rater styles, and a two-stage training strategy. Evaluation is performed on two public multi-annotator datasets, one for CT abdominal organ segmentation, the other for brain MRI, and against several existing methods for few-shot image segmentation.
- 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.
This work’s main strength is that it links the tasks of few-shot learning and multi-annotator supervision, a novel idea. Additionally, the authors claim that the proposed attention-based calibration module is lightweight and can easily be inserted into existing prototype-based models, an elegant and practical approach. Finally, the proposed method is evaluated using two public datasets of different modalities, CT and MRI, showing improvement to the state of the art for few-shot image segmentation.
- 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.Significance of results: Though the results in Tables 1 and 2 show improvement of methods when including the authors contributions, the improvement is small and standard deviation is not reported to evaluate whether the improvement is truly significant. The same holds true for the ablation study results in Table 3. 2.Pseudo-labelling: I am uncertain whether the pseudo-style generation, using morphological operations on super-pixels, truly reflects human annotation variability. 3.Multi-rater analysis: capturing multi-annotator style when training models is listed as one of the main contributions of this work. However, the authors do not show concrete results of how and if the trained models capture different annotation styles and biases. I would expect a deeper analysis of the multi-rater modeling.
- 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
Reproducibility: consider releasing code if accepted, it is not mentioned if this will be made public
Results:
- consider including deviation of mean to tables
- multi-rater modeling: since this is a key contribution of this work, I would expect more information on how the model is learning annotator styles and rater biases.
Fig 2: It is not clear to me in this image if the model is picking up annotator style. Consider another example here or an explanation. I would also like to see all three raters, the figure covering the entire row for easier visualization, and the title explaining which model predictions these correspond to with dice scores for this case etc.
Methods: ensure all variables in formulas are explained (e.g. H and W in 2.1)
Pseudo-style generation: clearer explanation of how these labels are used in training to improve model performance
- 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?
While as mentioned above I believe the results are not currently displaying a clear performance gain, I would accept this paper if the results are adjusted to display standard deviation and it is clear that the suggested method outperforms existing approaches, since I believe it is an innovative idea with a clear implementation in an unexplored field.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper introduces a novel setting of few-shot medical image segmentation. The authors propose an Attention-Based Prototype Calibration framework (JAPC), which refines rater-specific prototypes by modeling their deviations from a consensus prototype using an attention mechanism. Additional components include a calibration loss, pseudo-style generation to simulate rater variability, and a two-stage training strategy for stability. Experiments on CURVAS (CT) and QUBIQ (MRI) show consistent improvements over prototype-based few-shot 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.
The paper is well-motivated and addresses a realistic yet underexplored problem, namely handling multi-rater variability in few-shot segmentation, which is highly relevant for clinical practice. The proposed formulation is intuitive and meaningful, avoiding consensus collapse while enabling personalized predictions. The method is also lightweight and modular, as the prototype calibration operates directly in prototype space without modifying the backbone, making it broadly compatible with existing few-shot segmentation frameworks. Empirically, the approach demonstrates consistent improvements across datasets, settings, and raters, suggesting robustness and practical value. The inclusion of ablation studies further helps validate the contribution of each component.
- 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 evaluation is somewhat limited, as it only considers two relatively small-scale datasets and does not compare against strong uncertainty-aware or probabilistic multi-rater methods adapted to few-shot settings. It also lacks analysis of generalization to unseen rater styles and robustness to varying numbers of raters, making broader applicability unclear. In addition, the performance gains are modest (≈1–2%) despite increased model complexity, including multiple additional modules and training strategies. The improvements may not fully justify the added design overhead, and performance remains limited on more challenging structures such as the pancreas. Finally, there are minor clarity issues in notation, as some abbreviations (e.g., fg/bg) are not explicitly defined when first introduced, which slightly affects readability.
- 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?
While the paper addresses a relevant and practically important problem and demonstrates consistent empirical improvements, its overall contribution appears incremental. The proposed method combines several existing ideas in a reasonable but not fundamentally novel manner. The evaluation is also limited in scope, being restricted to two relatively small datasets, without comparison to stronger uncertainty-aware or probabilistic multi-rater baselines, nor does it examine generalisation to unseen rater configurations. Additionally, the reported performance gains are modest given the added architectural complexity, raising questions about the cost–benefit trade-off. Overall, the paper is technically sound and well-motivated, but the novelty and empirical validation are not sufficiently strong.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Though my concerns about limited evaluation and incremental novelty are only partially addressed – particularly the lack of improvement on challenging structures – I believe the paper is worth accepting. The authors convincingly demonstrate that JAPC adds minimal overhead while achieving statistically significant gains across multiple few-shot settings. They also provide missing multi-rater comparisons and show generalisation to different rater counts and unseen rater styles. The remaining limitations are acknowledged and can be addressed in future work.
Author Feedback
We thank reviewers for the comments and for recognizing JAPC’s motivation, novelty, and modularity. R2 calls it “well-motivated and sound,” R4 “novel,” “elegant and practical,” and R5 notes its “robustness and practical value.”
(1) Significance of gains, overhead (R2,R4,R5). Each test scan is used once as support, giving 65 CURVAS results. One-tailed paired p-tests comparing ours+base vs. base show all gains are significant, with p<0.01.For SSL-ALPNet/DSPNet, p-values are 5.25e-19/2.71e-24 in Setting 1 and 5.84e-12/2.20e-65 in Setting 2.JAPC adds only ~0.2% parameters to SSL-ALPNet and ~0.3% to DSPNet, and does not increase training iterations: it starts from the T/2 checkpoint and trains for the remaining T/2 iterations. Thus, JAPC yields consistent low-cost gains.
(2) Novelty and multi-rater comparison (R5). Our goal is scarce, heterogeneous supervision, not fully supervised uncertainty modeling. Existing multi-rater methods usually assume abundant labels and do not directly handle episodic few-shot adaptation to unseen classes (see paragraph 2, Sec. 1). Under few-shot constraints, we compare with strong multi-rater methods, MRNet and D-Persona trained on the support set in Setting 1 (Table 1). They cannot be used in Setting 2 because test classes are unseen during training.
(3) Multi-rater modeling evidence (R4). JAPC follows the observation that raters agree on the main object but differ near uncertain boundaries. The consensus prototype captures shared appearance; rater-specific prototype differences capture each rater’s shift from consensus. Calibration lets each rater “see” these shifts and make separate predictions instead of averaging raters. Fig. 2 supports this: predicted rater differences occur in regions similar to true label differences, and rater-wise Dice improves for all raters. We will release a qualitative grid.
(4) Generalization to rater configurations (R5). CURVAS has 3 raters and QUBIQ Brain-Growth has 7, showing JAPC is not tied to a fixed small rater count. If test-time rater count is unknown, consensus masks can be added as extra raters to match the attention matrix dimension. Real rater styles are unseen during training: training uses pseudo labels independent of real styles, while evaluation uses real rater annotations. Thus, results test generalization to unseen real styles.
(5) Ablation clarification (R2). In Table 3, when “Attn” is off, JAPC attention calibration is removed and rater prototypes pass through as-is.
(6) Scope of evaluation (R2,R5).
Architecture. SSL-ALPNet and DSPNet share a similar feature extractor but differ substantially in prototype modeling, covering foundational and recent SOTA pipelines. JAPC operates on intermediate representations, so it is model-agnostic and applicable to other backbones, including transformers.
3D generalization. It is also dimension-agnostic, since it uses features/prototypes rather than 2D-specific designs, and can in principle integrate into 3D volumetric or supervoxel frameworks.
Datasets. We evaluate on complementary CT and MRI datasets with different anatomy and multiple annotators. More/larger datasets and 3D experiments are future work. CURVAS is not small-scale: it has abdominal CT volumes with ~900 axial slices per patient, 3 organ classes, and 3 raters.
(7) Pseudo-style generation (R4). We apply random 1- or 2-step boundary transformations to each pseudo label: erosion, dilation, opening/closing, hole filling, small-component pruning, directional boundary shifts, and Gaussian blur, with small kernels to preserve semantics. They need not match each rater exactly; they teach that rater variability often appears as structured boundary shifts around a shared object. At test time, real support annotations provide rater-specific styles, and JAPC calibrates prototypes from those masks.
Minor issues. We appreciate the notation, figure, and code-release suggestions. If accepted, we will clarify notation and release code.
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 find the problem formulation of few-shot multi-rater segmentation to be clinically meaningful, and the proposed attention-based prototype calibration framework to be technically sound and well-motivated. However, concerns remain regarding the limited empirical validation and modest performance improvements. The evaluation is restricted to relatively small datasets without comparison to stronger multi-rater or uncertainty-aware baselines. Additionally, the paper does not sufficiently demonstrate how the model captures annotator-specific styles, and the realism of the pseudo-style generation process requires further justification. The rebuttal should focus on clarifying experimental settings and providing deeper insights of multi-rater modeling.
- 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 rebuttal satisfactorily clarifies statistical significance, ablation settings, overhead, pseudo-style generation, and multi-rater modeling. All reviewers updated to accept, including the initial weak reject. While gains are modest and broader validation would strengthen the work, the method is lightweight, technically sound, and supported by consistent evidence. I recommend 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.
While this submission received mixed scores during the initial stage, I believe that the rebuttal positively addressed the most important raised concerns. Therefore I recommend Accept for this work and strongly suggest the authors to integrate these clarifications/responses in the camera ready version of the paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Strengths: The reviewers generally recognize the clinical significance and clear motivation of formulating the few-shot multi-rater segmentation problem. The proposed Joint Attention-based Prototype Calibration (JAPC) framework is technically sound, elegant, and modular. The approach of decoupling consensus appearance from rater-specific style shifts represents a well-motivated and promising idea for dealing with label ambiguity under scarce supervision constraints. Weaknesses and Areas for Improvement: Despite the authors’ detailed rebuttal and the interesting core idea, significant concerns remain regarding the empirical validation, baseline comparisons, and the realism of the methodology, which prevent the paper from acceptance in its current form: 1.Insufficient Baseline Comparison & Empirical Validation: The evaluation remains a major weakness. While the authors clarified that existing fully-supervised multi-rater methods are difficult to adapt to episodic few-shot scenarios, the comparison with SOTA methods is overly restricted. The lack of comparison with stronger, more diverse multi-rater or uncertainty-aware baselines leaves the true competitiveness of the proposed method unverified. Furthermore, the evaluation is conducted on relatively limited datasets. 2.Over-reliance on Marginal Improvement Signatures: The authors argue the significance of their performance gains using one-tailed paired p-tests. owever, in the context of clinical image segmentation, a statistically low p-value derived from a large number of slices (e.g., in CURVAS) does not automatically guarantee practical or clinical significance, especially when the absolute performance improvements remain modest. 3.Justification of Pseudo-Style Generation Realism: While the authors explained that the pseudo-style generation aims to teach the network structured boundary shifts, the realism of this generation process remains questionable. Simple geometric or morphological perturbations do not sufficiently mimic the complex, expertise-driven, and subjective variations exhibited by real clinical annotators. The gap between heuristic pseudo-styles and real-world clinical rater profiles requires deeper justification or a data-driven validation. Overall, this paper introduces a promising idea with solid motivation. However, due to the limited scope of experimental validation, the modest absolute gains, and the unverified realism of the pseudo-style simulation, I agree that the paper requires a more thorough round of revision and stronger experimental backing. Therefore, a rejection is recommended. The authors are highly encouraged to incorporate the feedback—particularly expanding the baseline comparisons and refining the style modeling—for a future submission.