Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Segment Anything Models (SAM) enable prompt-driven medical image segmentation with minimal annotation. Propagation-based methods extend this paradigm to volumetric data by iteratively propagating semantic guidance from a single prompted slice. However, existing approaches suffer from two critical limitations: error accumulation, where early prediction errors compound, and fixed propagation strategies that fail to adapt to heterogeneous anatomical structures. To address these issues, we propose a training-free Reliability-Aware Cross-Prompt Aggregation framework. Instead of relying on a single evolving prompt, our method jointly leverages two complementary prompts at each propagation step: the initial prompted slice providing stable global guidance, and the dynamically propagated front slice capturing local appearance. Their predictions are adaptively fused through a reliability-aware mechanism that combines spatial distance weighting with an unsupervised estimation of prompt reliability. This allows the model to intelligently down-weight unreliable predictions and prevent error compounding. Experiments on ten external multi-modal medical imaging datasets demonstrate that our approach consistently outperforms state-of-the-art Medical SAM baselines across all six evaluation metrics, achieving an average Dice score improvement of 4.6% over the best baseline.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1506_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiHao_ReliabilityAware_MICCAI2026,
        author = { Li, Haoshen AND Chen, Zifan AND Yang, Guangzhengao AND Chen, Guanyu AND Lin, Ziling AND Nan, Xinyu AND Zhao, Jie AND Zhang, Li},
        title = { { Reliability-Aware Cross-Prompt Aggregation for Propagation-Based Segmentation of Arbitrary 3D Medical Objects } },
        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 introduces a training-free extension of the prompt propagation PAM paradigm for achieving volume-aware 3D segmentation using an underlying 2D-only model. The extension is based on the fusion of two predictions: from the original prompt and the propagated front prompt. The fusion is weighted by three metrics: distance, cycle consistency, and prediction agreement.

  • 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 presents the novel idea of fusing the initial prompt with the front prompt for volume-aware segmentation. The use of cycle consistency as a measure of prediction reliability is a particularly elegant contribution. The reported Dice gains appear high in absolute terms, although the results exhibit considerable variability, as evidenced by high standard deviations.

  • 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 has two major weaknesses. First, the computational cost of the proposed method appears to be substantial (if I understood correctly), as it requires one additional inference to obtain both the initial and front prompt predictions, plus two further inferences for computing cycle consistency; this overhead is not discussed anywhere in the paper. Second, the lack of genuine external validation is a concern: the pretrained PAM model used was trained on the same datasets employed in this study, so no true out-of-distribution evaluation is provided. As a minor observation, the ablation studies reveal that performance is already quite high with the initial prompts alone (96.7% of best performance), while distance-based fusion reaches 99.3%; this raises the question of whether the additional inference steps, and those required by cycle consistency in particular, are truly justified. Finally, the proposed methodology has a somewhat limited novelty with regards to the PAM approach.

  • 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

    The cycle consistency and cross-prompt agreement metrics appear to be most informative in a homogeneous settings with minimal deviation between the front and initial prompts; a demonstration on more extreme cases, involving large geometric or textural differences, would help establish the robustness of these metrics. Additionally, it is unclear whether the standard deviation reported in Table 1 is computed across the 10 datasets or across the combined pool of samples — this should be clarified. Finally, the radar charts are somewhat difficult to read due to the shading; removing it would help reduce visual clutter.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The somewhat limited novelty of the proposed methodology would put this paper at a Weak Accept, but the lack of discussion on computational resources and failure to provide a genuine external validation severely dimish the quality of this paper.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    My two main concerns were the computational cost and the issue with proper external evaluation.

    The computational cost is somewhat addressed in the rebuttal: while the authors don’t provide a full analysis, they report wallclock times for their method. Even though the newly proposed method adds significant overhead (almost 3x compared to PAM), the wallclock time appears to be in line with other SOTA methods, while achieving better performance.

    For the lack of true external validation, it seems I had misunderstood the fact that only the testing datasets from the PAM paper had been used in this new work as well. I would ask the authors to clarify this in the manuscript, but in conclusion I apologize for the misunderstanding on my part.

    The other remaining concerns are the somewhat limited novelty, and whether the performance of the method justifies its complexity, in particular in relationship to simpler versions in the ablation studies. For the novelty, this is a decision for the meta-reviewer and area chair to judge whether this constitutes a reason for rejection. The complexity comment is not fully addressed in the rebuttal: the authors report very-strongly-significant p-values as a justification, but my comment was more focused on the tradeoff of the computational cost vs obtained effect. To be clear: the distance-only ablation (which if I understood correctly only requires minimal additional computational cost because of lack of cycle consistency) achieves 0.006 DSC less than the full model. Is the computational cost of the full model justified to achieve 0.006 DSC better? Overall, in my opinion, the rebuttal improved the quality of the paper to marginally below the threshold for acceptance, I do not feel strongly about rejection.



Review #2

  • Please describe the contribution of the paper

    Training-free extension of PAM for propagation-based 3D medical segmentation. Fuses predictions from the fixed initial prompt and a dynamically propagated front slice via a distance-aware exponential weight and a confidence-aware weight built from two unsupervised reliability metrics: cycle consistency and cross-prompt agreement. +4.6% mean Dice over PAM across 10 external multi-modal 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.
    • Training-free, plug-and-play on top of a frozen PAM — easy to adopt in the foundation-model era.
    • Problem diagnosis is sharp: error accumulation and rigid distance-based refresh are correctly identified, and the two-prompt design directly addresses both.
    • Strong empirical breadth — 10 external datasets across CT, MR, and SRX, six metrics, consistent improvements across the board.
    • Ablation cleanly isolates each component: single-prompt (0.803) → average fusion (0.820) → distance-only (0.825) → confidence-only (0.826) → full (0.831), so every contribution is justified.
    • The reliability metrics are validated empirically — Pearson ≈0.63, Spearman ≈0.69 correlation with true Dice — which is a rare and useful piece of analysis.
  • 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.
    • Novelty is incremental: cycle consistency and cross-modal/prompt agreement are well-known ideas from SSL and interactive-segmentation work; the contribution is a careful combination rather than a new mechanism.
    • No runtime or FLOPs analysis, which is critical for a training-free method — cycle consistency requires backward propagation to both prompts at every slice (roughly 3× the vanilla PAM cost).
    • τ = 100 is hand-picked with no sensitivity analysis, yet slice spacing varies substantially across the 10 datasets.
    • Initial-prompt confidence is hard-coded to 1; it is never down-weighted even when propagation moves far from it.
    • Missing baselines: no SAM2, no nnU-Net reference upper bound, and no simple ensembling baseline (e.g., averaging propagation restarts from different initial slices).
    • No statistical significance testing on per-dataset differences; some radar-chart gains may lie within variance.
    • Writing is serviceable but generic; a sharper conceptual argument for why cycle consistency and cross-prompt agreement are complementary (not redundant) would strengthen the paper.
  • 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
    • What is the inference-time cost per volume compared to vanilla PAM (wall-clock and GPU memory)?
    • Does τ need to be retuned per dataset/modality? Please show a sensitivity curve.
    • Are per-dataset improvements over PAM statistically significant (paired Wilcoxon)?
    • What happens if the initial-prompt confidence is also computed from the same two metrics instead of being fixed at 1?
    • How much of the 4.6% Dice gain is captured by a simpler baseline of averaging the two prompt predictions (0.820 in Table 2) plus distance weighting only, vs. the full model?
    • Are cycle consistency and cross-prompt agreement themselves correlated? If so, how complementary are they really?
  • 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 method is training-free, well-motivated, broadly evaluated (10 external datasets × 6 metrics), and consistently better than SegVol, MedSAM, MedSAM2, and PAM. The ablation cleanly isolates each component and the correlation study gives empirical evidence that the unsupervised reliability metrics actually track Dice. The remaining reservations — incremental novelty, missing runtime analysis, no τ sensitivity study, and no statistical testing — are all camera-ready-fixable rather than fundamental. Net positive for the community.

  • 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

    To addresses the two limitations of propagation-based 3D medical image segmentation methods: error accumulation over long-range propagation and fixed propagation strategies that fail to adapt to heterogeneous anatomical structures, this paper proposes a training-free, reliability-aware cross-prompt aggregation framework. The contributions are as follows: novel cross-prompt aggregation strategy, unsupervised reliability estimation, and distance-aware weighting.

  • 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 addresses error accumulation without training, using both global (initial) and local (front) prompts. 2.This paper adaptively weights prompt predictions via two unsupervised metrics: cycle consistency and cross-prompt agreement. 3.Experiments on ten external multi-modal datasets show consistent improvements across six metrics over the SOTA methods.

  • 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.This paper requires computing forward passes per slice plus cycle consistency evaluations. This maybe doubles inference time compared to vanilla PAM. The paper does not report inference time or computational cost. 2.The weight τ is fixed. The paper does not present sensitivity analysis for this hyperparameter across different slice spacings or modalities.

  • 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.

    (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?

    novel cross-prompt aggregation strategy, impressive experimental results

  • 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.

    The rebuttal resolves my main concerns, so my decision is Accept.



Author Feedback

We thank all reviewers for their comments and for recognizing the clear motivation, novel/well-suited method and thorough evaluation. We address main concerns below.

Q1: Computational cost (R1-R3) We report per-volume inference time on HaN-Seg, which has the most objects (1259) and organs among ten datasets. On one NVIDIA A800-SXM4-80GB GPU, the times are: SegVol 8.08s, MedSAM 8.26s, MedSAM2 9.60s, PAM 2.86s, and Ours 7.11s. Although our method adds inference overhead compared to vanilla PAM, it remains faster than the other methods and improves mean Dice by 4.6% on ten external datasets, showing a favorable accuracy-efficiency trade-off.

Q2: External validation, baselines, and statistics (R1,R2) All ten reported datasets are external and unseen during PAM training. Following original PAM setting, PAM was trained/validated on 34 internal datasets and additionally evaluated on ten external datasets. Our method uses the pretrained PAM weights and reports only external results, demonstrating out-of-distribution generalization. For baselines, MedSAM2 represents the medical extension of SAM2 and is included. We added supervised nnUNet on Adrenal-ACC-Ki67-Seg and ACDC. Dice scores are: nnUNet 0.854/0.929, PAM 0.891/0.860, and Ours 0.908/0.883.Our method remains competitive with nnUNet and can be better when annotations are scarce or targets are complex. For statistics, paired Wilcoxon tests on Dice show that our method significantly improves over PAM on all ten datasets (all per-dataset p<1e−5). To address R1’s concern about whether extra inference steps are justified, we report p-value on the merged ten dataset: Ours vs. initial-prompt only, p=4.41e−92; Ours vs. distance-only, p=5.34e−04, confirming the value of cross-prompt inference and reliability-aware weighting.

Q3: Sensitivity to tau and initial-prompt confidence (R2,R3) Because distance weighting uses physical distance (slice interval multiplied by spacing), tau is less sensitive to spacing variations. We further conducted a sensitivity analysis with tau in {50, 100, 200}, and mean Dice is 0.828, 0.831, 0.829, within a narrow 0.003 range, showing robustness. The slight drops at tau=50/200 are expected: smaller tau over-emphasizes the front slice, while larger tau weakens distance discrimination and makes weights closer to uniform. We set initial-prompt confidence to 1 due to its high reliability and lack of reference slice for reliability estimation. Still, when propagation moves away, its distance weight decays, reducing total contribution.

Q4: Novelty (R1,R2) Unlike PAM’s single-prompt fixed propagation, we propose cross-prompt aggregation with two novel training-free reliability estimation metrics for adaptive prompt fusion. Cycle Consistency and Cross-Prompt Agreement estimate reliability without labels and guide adaptive fusion of initial/front predictions, suppressing unreliable predictions and reducing error accumulation. Their positive correlations with true slice Dice also suggest value for future segmentation studies.

Q5: Other issues of R1 (1) Robustness in challenging cases: ACDC cardiac MRI exhibits large basal-to-apical geometric variations. As shown in Q2, our method significantly improves over PAM on ACDC, showing robustness. (2) Standard deviation in Table 1 is computed across the ten datasets. (3) We will remove radar-chart shading to improve readability.

Q6: Other issues of R2 (1) In Table 2, “Average” is equal-weight two-prompt fusion, while “Distance only” uses distance-aware weights without confidence metrics. Thus R2’s simpler baseline (average plus distance weighting) equals our “Distance only” setting, with Dice=0.825.(2) Cycle Consistency and Cross-Prompt Agreement are positively correlated (average Pearson/Spearman=0.6875/0.6810), as both reflect reliability, but are not redundant: the former measures backward self-consistency, while the latter measures two-prompt consensus. Correlation below 1 indicates related but complementary cues.



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 work proposes a reliability-aware cross-prompt aggregation method for propagation-based 3D medical image segmentation. The recommendation is mixed among reviewers. The main disagreements are on methodology novelty, empirical evaluation and lack of sufficient evidences and explanations for some claims. Based on my reading, I think that the reviewers raised effective concerns. Please try to fix these concerns in the rebuttal.

  • 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.

    This paper received very mixed ratings, one for accept, one for reject and one for weak accept but not read the rebuttal file. I read the paper and rebuttal file, and think that the authors solve the main concerns though some important issues still need to be fixed. But due to the novelty of the method, I tend to accept.



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.

    NashNet introduces an innovative, training-free wrapper that elegantly addresses error compounding in volumetric slice propagation by implementing an unsupervised dual-prompt aggregation architecture. The authors successfully resolved initial concerns surrounding computational overhead and external validation during the rebuttal, demonstrating that their method remains highly competitive with established baselines like MedSAM2 while achieving a 4.6% mean Dice increase across ten completely unseen external datasets. Highly significant statistical verification ($p < 1e-3$) proves the value of its cycle-consistency loop over simpler distance-based baselines, this work represents a robust and clinically viable contribution.



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.

    This submission is a borderline case, where several of the concerns are not fully resolved. After reading the paper, I feel that the novelty of the proposed work is rather incremental, and that several of the proposed modules are not actually required. For instance, as highlighted by one reviewer, the trade-off of including more components (beyond the distance aware scheme) is questionable, as it brings negligible gains compared to the additional computational cost, which may undermine even more the actual contribution. Furthermore, while overall gains seem to improve over the baseline PAM, per-dataset scores are not reported, making it hard to assess whether the proposed PAM extension is indeed a consistently better approach, or it largely outperforms other methods in a small number of datasets, falling behind in the remaining ones.

    Given these concerns, and that there are not strongly supportive comments from the reviewers (beyond the large number of datasets in the empirical validation), I am inclined towards recommending rejection for this work.



back to top