List of Papers Browse by Subject Areas Author List
Abstract
Parameter-Efficient Fine-Tuning (PEFT) enables foundation models like the Segment Anything Model (SAM) to be adapted for medical image segmentation with limited supervision. However, limited supervision from a single domain tends to collapse the inherent generalization capacity of SAM toward source domain-specific styles, leading to performance drops on unseen clinical centers. To address this, we propose Symmetric Spectral Uncertainty Gating (SSUG) for SAM-based Single-Domain Generalization (SDG), which explicitly disentangles anatomical representation from styles through a synergistic closed-loop design. First, a Spectral Manifold Extrapolation (SME) module proactively probes the statistical boundaries of the source spectrum to simulate various possible stylistic shifts. Second, an Uncertainty-aware Symmetric Channel Gating (USCG) module measures response divergence to softly calibrate style-sensitive channels. Finally, we introduce Hierarchical Decoupled Consistency (HDC) constraints to reinforce domain-invariant anatomical representations across multiple feature granularities. Experimental results on two challenging domain generalization benchmarks for Fundus and Prostate segmentation proved the significant superiority of SSUG compared to state-of-the-art SDG methods. Code is available at https://github.com/HiLab-git/SSUG.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4696_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HiLab-git/SSUG
Link to the Dataset(s)
Fundus dataset: https://github.com/emma-sjwang/Dofe
Prostate dataset: https://github.com/liuquande/SAML
BibTex
@InProceedings{LiHe_Segment_MICCAI2026,
author = { Li, He AND Gu, Ran AND Fu, Jia AND Ye, Ping AND Wang, Litingyu AND Li, Kang AND Wang, Guotai AND Zhang, Shaoting},
title = { { Segment Anything Model with Symmetric Spectral Uncertainty Gating for Generalizable 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
The paper proposes a novel Symmetric Spectral Uncertainty Gating to adapt the Segment Anything Model for single-domain generalization in medical image segmentation. By mitigating style sensitivity and acquisition-based biases, it disentangles anatomical representations from domain-specific styles. Experimental results on two medical image segmentation benchmarks show the proposed method successfully improves the model’s cross-domain generalizability.
- 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) Strong clinical motivation and relevance: The paper targets the issue in real-world clinical practice: domain shifts caused by variations in data acquisition protocols and equipment-based biases. By explicitly addressing how these acquisition-based variations degrade the performance of foundation models in single-domain generalization settings, the authors provide a clear, clinically valuable motivation for their proposed framework. 2) Plug-and-play versatility: The proposed uncertainty-aware symmetric channel gating module is a lightweight plug-in component used in SAM. This module can be readily utilized in other models featuring Transformer mechanisms, which significantly broadens the potential impact and applicability of this work beyond just SAM. 3) Comprehensive experimental design: The evaluation is strong and convincing. The authors thoroughly validate their method across completely different anatomical regions and imaging modalities (Fundus 2D images and Prostate 3D MRIs) using multiple generalization datasets. Furthermore, the evaluation metrics are standard and reasonable, the selected baselines are both abundant and up-to-date (including recent SAM-based SDG methods), and the ablation studies prove the effectiveness of each individual component (SME, USCG, and HDC).
- 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 assumption about feature channels is too simple: The method assumes we can simply turn off “style” channels without losing “anatomy” information. However, in deep neural networks, a single channel usually contains both. If we completely turn off a channel to remove noise, we might accidentally delete important medical details (like small lesions). This is risky for clinical diagnosis. The authors should discuss this problem. 2) Potential high computational cost for real clinical use: The paper claims the module is “lightweight,” but the overall training process uses a dual-branch design and Fourier transforms. Since SAM is already a very large model, running two branches at the same time will cost a huge amount of memory and computing power. The paper needs to provide an analysis of the training time and memory usage to show if real hospitals can actually run this model. 3)Sub-optimal performance on specific metrics and missing failure analysis: Although the method achieves the best overall average performance, it does not get the best results on certain sub-tasks. For example, in Fundus Disc segmentation (Domain 2), The proposed method’s ASD score is worse than existing methods like MaxStyle and SLAug. The paper does not explain why the framework struggles in these specific situations. Furthermore, even though the table caption briefly mentions a significance test (p-value < 0.05), the text does not provide details on the exact statistical method used (for example, which specific test was applied to compare the models).
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
Although there are some limitations regarding the methodological hypothesis and real-world clinical deployment, this paper brings much-needed attention to the important but easily neglected issue of varying imaging protocols and acquisition-based biases. It provides a valuable new perspective on tackling domain shift in foundation models.
- 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.
Thank the authors for the rebuttal. The clarification on soft channel reweighting, training-only SME branch, inference cost, SME band setting, and statistical testing addresses most of my concerns.
My remaining minor concern is that the explanation for the worse ASD in some domains is still limited. I encourage the authors to add representative failure cases or case-wise discussion in the final version.
Review #2
- Please describe the contribution of the paper
1)A Spectral Manifold Extrapolation module is proposed to proactively probe the statistical boundaries of the source domain. 2)An Uncertainty-aware Symmetric Channel Gating module is introduced to quantify the sensitivity of each channel to style variations and dynamically suppress style-correlated channels accordingly. 3)Hierarchical Decoupled Consistency constraints are designed to reinforce domain-invariant anatomical representations across multiple feature granularities.
- 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.
It addresses the issue of generalization degradation of SAM in medical image segmentation under the single-domain generalization andfocuses on mitigating style sensitivity and overfitting within SDG foundation model architectures.
- 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 integration of the method description and the research motivation is relatively weak, such as the summarized modules and the issue of style sensitivity in SDG is not directly evident in the Introduction, and the experimental section lacks sufficient analysis of the results.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 method demonstrates clear novelty and the experimental results are sufficient to provide valuable insights for related research fields.
- 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.
The question addressed in this paper is novel. In the rebuttal, the authors provide reasonable explanations for the effectiveness and novelty of the method, and appropriately address the parts of the original manuscript that might have been misunderstood. I recommend accepting the paper.
Review #3
- Please describe the contribution of the paper
This paper aims to address source-domain overfitting in single-domain generalization for medical image segmentation using a SAM-based framework. The framework incorporates modules such as SEM, USCG, and HDC.
- 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 three modules form a coherent and well-structured framework: SME simulates style variations, USCG identifies and filters style-sensitive channels, and HDC enforces consistency between the original and augmented images. 2.The paper compares traditional data augmentation methods with state-of-the-art domain generalization approaches for SAM, demonstrating superior performance on two challenging datasets and providing compelling empirical evidence.
- 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 novelty of the approach appears limited. Amplitude perturbation in the frequency domain for data augmentation is a well-established technique in medical image domain generalization. The method primarily combines this operation with exponential moving average (EMA) for statistical extrapolation. USCG and HDC can be interpreted as variants of attention mechanisms and feature consistency regularization, which are commonly used in multi-view contrastive learning. The integration of these existing techniques within the SAM framework appears to constitute incremental engineering-level innovation. 2.SAM is inherently a prompt-driven model. It is unclear how prompts are obtained during the testing phase. Specifically, it is not specified whether ground-truth bounding boxes are used or whether the model operates without prompts. The paper only states that the prompt encoder is “fully fine-tuned with default embeddings,” which remains insufficiently specified. Given that SAM’s performance is highly sensitive to prompt design, this lack of detail raises concerns about the fairness and validity of the quantitative comparisons. 3.SME reconstructs the image within a band-limited region; however, the definition of this region is not clearly specified.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
This assessment is primarily based on several unclear experimental settings; however, the evaluation could be reconsidered if the aforementioned issues are adequately clarified.
- 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
We thank the meta-reviewer and reviewers for their constructive comments and for recognizing the clinical relevance (R2), coherent design (R4/R5), and solid empirical results (R2/R5) of our work. We clarify the main concerns below:
*Prompt configuration and fair comparison(R5) In our method, we aim at automatic segmentation of medical images, and SAM was used with its everything mode, so no human interaction or Ground-Truth-derived prompt (bounding box, point, or text prompt) is needed. This ensures a fair comparison with the competing methods that do not use human interaction as well.
*Soft uncertainty calibration, not hard channel removal(R2) We agree that a channel may contain both anatomical and style-related information, and do not assume a binary separation of “style” and “anatomy” channels. Instead, we quantify each channel’s sensitivity to style change (Eq.2) and use soft and progressive reweighting to calibrate each channel (Eq.3), rather than hard deletion. HDC further constrains both calibrated intermediate features and final embeddings, encouraging anatomical priors to remain robust throughout the entire encoding process.
*Systematic novelty beyond simple integration(R5) We acknowledge that spectral perturbation, gating, and consistency regularization have been studied individually. However, S²UG is not a simple aggregation of generic techniques; it is a synergistic closed-loop system tailored to counteract source-style overfitting in PEFT-adapted SAM under SDG. Specifically, SME is not a passive or random amplitude perturbation: it estimates source spectral statistics in log-radial space and extrapolates along the statistical envelope to actively probe distribution boundaries (Eq.1). USCG is also not generic channel attention; it is an uncertainty-aware gate explicitly conditioned on the original/SME-branch response divergence for symmetric channel calibration inside the SAM encoder. Together with HDC, these modules form a coupled cycle of active style probing, sensitivity calibration, and anatomy-preserving regularization, rather than a simple combination of existing modules.
*Computational cost and deployment efficiency(R2) Note that FFT/IFFT-based SME and the augmented branch with consistency regularization are only needed for training. As a result, our method does not increase computation or memory cost for inference: only the original image is processed through the standard branch with the learned USCG modules, and no manual/GT prompts are needed. In Fundus inference, S²UG runs at about 0.35s/case with the GPU memory footprint of 1.87 GiB, which makes it easy for clinical deployment.
*Implementation details and statistical clarification(R2&R5) In SME, the band-limited region is the low-to-mid FFT amplitude band with normalized radius ρ∈[0.25,0.45], where amplitudes are extrapolated using the estimated spectral envelope while out-of-band amplitudes and the original phase are preserved. Statistical significance was computed using a two-sided paired t-test on per-case Dice/ASD scores between S²UG and the second-place method, with p<0.05 considered significant, which will be clarified in the final version with minor modification.
*Sub-metric analysis and overall robustness(R2&R4) Due to the complexity of segmentation objects and inter-domain variation, there is no single method that always achieves the best value on all evaluation metrics in all domains on all datasets. S²UG achieves the best average Dice/ASD values on both datasets, demonstrating its general robustness. In some domains, such as Domain 2 in Table 1, the relatively large ASD is mainly due to the fact that ASD is sensitive to noise (false positives far from the foreground). We will clarify this limitation in the revised manuscript and provide a more detailed case-wise discussion in a future journal extension.
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 paper addresses an important problem (single-domain generalization for SAM in medical imaging) and shows solid empirical results across multiple modalities. Two reviewers assign Weak Accept (4), one assigns Weak Reject (3). The main concerns are: (1) unclear SAM prompt configuration during testing, which is critical for fair comparison; (2) the novelty is primarily integrative, combining existing techniques within the SAM framework; (3) missing computational cost analysis; and (4) oversimplified assumption about style/anatomy channel disentanglement. These concerns are addressable during rebuttal, though the prompt configuration issue is critical and must be clarified. The paper should proceed to rebuttal with a clear expectation that the authors address these points.
- 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 presents a well-motivated S²UG framework to improve SAM’s single-domain generalization via spectral uncertainty gating, with solid experiments on fundus and prostate MRI. Reviewer #2 and #4 gave Weak Accept (4), and Reviewer #5’s Weak Reject (3) concerns (prompt configuration, novelty, computational cost) were adequately addressed in the rebuttal by clarifying the automatic prompt-free mode, soft channel calibration, and inference efficiency (0.35s/case, 1.87 GiB). The overall positive consensus supports 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.
This paper addresses an important and timely problem: single-domain generalization for SAM-based medical image segmentation. The proposed S2UG framework is coherent: SME actively probes spectral style shifts, USCG performs soft response-divergence-based channel calibration, and HDC enforces multi-level consistency. The experiments cover two distinct benchmarks, include recent SAM-based SDG baselines, report Dice/ASD with statistical testing, and show consistent average improvements. The rebuttal satisfactorily resolves the most critical concern by clarifying that SAM is used in automatic “everything” mode with no GT-derived prompts, ensuring fair comparison. It also clarifies that USCG performs soft calibration rather than hard channel removal, that SME is training-only, and specifies the spectral band and statistical test. While novelty is partly integrative and some sub-metric failures deserve more discussion, the technical design and empirical evidence are sufficient for acceptance.
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.
I have read the paper, reviews (2 positive, 1 negative), and the rebuttal. Despite the positive scores, critical flaws remain. The evaluation lacks comparisons with traditional (non-SAM) domain generalization methods, making it impossible to validate the clinical viability of the results. Furthermore, the core FFT-based frequency method lacks novelty (extensively explored in prior work like FedDG), and the proposed architecture appears to be an ad-hoc, unmotivated combination of existing modules.
