List of Papers Browse by Subject Areas Author List
Abstract
Segmentation foundation models such as SAM2 generalize well to natural images, yet adapting them to heterogeneous medical modalities remains challenging due to severe domain shifts and scarce annotations. We propose BayesPrompt, a Bayesian prompt adaptation framework for few-shot cross-modality medical segmentation. BayesPrompt combines Bayesian Meta-Prior Adaptation (BMPA), which regularizes lightweight decoder updates via source-target posterior alignment, with a Probabilistic Prompt Module (PPM) that encodes class-wise feature statistics and predictive uncertainty into attention-conditioned prompt tokens. This design supports both gradient-free fast prompting and efficient Bayesian fine-tuning while mitigating overfitting and improving calibration. We evaluate BayesPrompt on ultrasound and MRI rotator cuff tear segmentation under single- and cross-modality transfer with k in {1,3,5,10} labeled target samples. BayesPrompt consistently improves Dice performance and robustness, demonstrating data-efficient adaptation of segmentation foundation models for medical imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6571_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Macs-Laboratory/bayesprompt
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HonSak_UncertaintyAware_MICCAI2026,
author = { Hong, Sakang AND Kim, Jun-Young AND Lee, Kyungsu},
title = { { Uncertainty-Aware Bayesian Prompt Adaptation for Robust Cross-Modality Medical 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
The work is about adapting segmentation models trained in a source domain to a target domain in a few shot training paradigm for the task of medical image segmentation. They take the SAM2 segmentation model and freeze its encoder weights but keep the decoder trainable for the medical image segmentation task. They propose a Bayesian Meta-Prior Adaptation module which they use to train on the few samples from the target domain by aligning the target posterior from the lightweight decoder/head parameters with a source-derived prior using the ELBO loss. This is pretty standard. However the part that is novel IMO is the Probabilistic Prompt Module where the authors generate prompts using the class masks for the image in the target domain and use these (as key and value) to compute attention (with the source image encoding) in the SAM2 decoder. Experiments are exhaustive with ablation studies included.
- 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 problem is interesting and relevant to clinicians/researchers who are building models in the domain of medical image segmentation. The paper is well-written and the evaluation has been done well with several SOTA models compared against. Ablation study has also been performed. The methodology is novel and has been well described.
- 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.
- Frozen backbone assumes SAM2 generalizes well across modalities, but severe shifts (US speckle vs. MRI smoothness) may still hurt deep features.
- k=1 to 10, but clinical deployment often 0-shot or test-time only. No zero-shot mode beyond vanilla SAM2.
- Uses only labeled supports—ignores abundant unlabeled targets for self-training/contrastive alignment.
- 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
Would be interesting to see how to use unlabeled samples to improve the performance of these models.
- 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?
Well written paper. Novel method. Relevant problem for researchers in the domain. Exhaustive evaluation.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes a framework for few shot and cross-modality segmentation. The main contribution is the combination of 3 modules: a regularizer that balances source and target data contribution, a probabilistic module that allows the model to modulate attention based on confidence on the target data, and an adaptation module that allows to balance between speed and precision. The paper also provides a comprehensive evaluation and comparison to other methods which demonstrate strong results on public and closed 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 major strength of the paper is the evaluation which has been done comprehensively including ablation and stability studies on both public and closed data. The results overall are very good compared to other state of the art methods with partial strong outperformance. Evaluation demonstrates that the method will likely generalize to other problems and application areas.
- 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 authors discuss reliability and stability in a separate section which is good. They also share that the target adaptation is limited to 20 iterations to avoid overfitting. I find this deserves more explanation and evaluation as it is quite common for these regularization terms to be very application specific and requiring trial and error to find the best number of epochs vs performance. This is a potential limitation of the work which should be addressed in more detail.
- 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
-
it would be better to report numbers for public datasets AMOS and BraTS separately for better comparability to other methods. I understand there are space constraints, but I find Fig. 2 could be optimised (see also below) to fee up space.
-
the method section does not contain any references. It would be good to cite closest related work, e.g. for the Gaussian meta prior in the regularisation term.
-
Figure 2, qualitative analysis: the segmentations are very small compared to the empty space around which makes it hard to compare against each other. It is also not obvious what structure in the CT/MRI is being segmented. It would be better to provide stronger zoom ins and maybe an overlay of the ground truth (boundary) with the original images.
-
Page 6.Fig X is not referencing any figure
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The evaluation shows strong performance in terms of segmentation accuracy and sufficient evidence has been presented regarding stability and generalizability.
- 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 authors propose BayesPrompt, a Bayesian prompt-adaptation framework for few-shot cross-modality medical image segmentation built upon SAM2.The method integrates uncertainty-aware prompting with lightweight parameter adaptation to improve generalization under limited supervision. The main contributions are a) a posterior-regularized adaptation mechanism that aligns the target distribution of lightweight decoder/head parameters with a source-derived meta-prior, aiming to stabilize few-shot learning and mitigate overfitting, b) a prompt generation module that aggregates class-wise feature statistics and predictive uncertainty from the support set to produce attention-conditioned prompt tokens, enabling modality-aware adaptation of the segmentation model and c) a flexible framework that supports both gradient-free prompt-based inference and lightweight Bayesian fine-tuning, providing a practical trade-off between efficiency and performance. Overall, the approach combines Bayesian-inspired regularization and prompt-based conditioning to address cross-modality adaptation in low-data regimes.
- 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.
- Consistent improvements over strong baselines across all k-shot settings, with the largest gains in low-shot regimes (k= 1,3).
- Achieves the lowest expected calibration error for the k=3-shot regime across compared 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)The evaluation is difficult to interpret due to inconsistencies between the baselines described in the experimental setup and those reported in the tables, as well as unclear grouping of methods across different experimental settings. Several methods mentioned as comparisons (e.g., VP-SFDA and MAUP, SAM2 zero-shot) are not clearly or consistently presented in the tables, while other methods (e.g.,CDUN at Table 1 , Prompt-Only at Table 2) appear without sufficient explanation or contextualization. In addition, the rationale behind splitting baselines across multiple tables is not clearly explained, and the relationship between different categories of methods (e.g., general segmentation models, SAM-based adaptations, and domain adaptation approaches) is not systematically organized. As a result, it is difficult to truly compare the results across all methods and fully assess the impact of the proposed approach in each context.
2) Minor presentation issues (e.g., unresolved figure references such as “Fig. X”) further affect readability.
3) In addition, the authors should explicitly highlight how the proposed framework outperforms competing methods while operating under the same or comparable SAM2 configuration, in order to clearly demonstrate the effectiveness of the approach under similar efficiency constraints
- 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 presents a well-motivated approach to few-shot cross-modality medical image segmentation by integrating Bayesian-inspired regularization with prompt-based adaptation on top of SAM2.The method addresses an important and practical challenge in medical imaging, namely achieving robust generalization under severe modality shifts while avoiding overfitting in low-data regimes.Empirically, the approach demonstrates consistent improvements over strong baselines, particularly in low-shot settings, and is supported by thorough ablation studies that help justify the contribution of its individual components. The incorporation of uncertainty-aware prompting is also a meaningful addition, especially in the context of reliability and calibration, which are critical for clinical applications. However, the evaluation suffers from clarity and organization issues that make it difficult to fully assess the method’s relative advantages across different settings. In particular, inconsistencies between described and reported baselines, unclear grouping of methods, and insufficient explanation of certain comparisons reduce the interpretability and reproducibility of the results. These issues appear to be primarily related to presentation rather than fundamental flaws in the methodology. Overall, the paper proposes a promising and practically relevant approach with solid empirical performance, but would benefit from improved experimental clarity and more structured evaluation.
- 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.
The paper is accepted because it presents a well-motivated and novel methodology with strong empirical validation, including comprehensive ablation and stability studies on both public and private datasets.
The reviewers noted consistent improvements over strong baselines across all k-shot settings, with particularly significant gains in low-shot regimes and improved calibration performance.
Overall, the work is clearly written, technically sound, and demonstrates strong potential for generalization and impact in medical image segmentation applications.