List of Papers Browse by Subject Areas Author List
Abstract
Single-Domain Generalization (SDG) in 3D medical image segmentation is inherently unstable: naive fine-tuning entangles domain-specific textures with anatomical semantics, leading to overfitting and catastrophic forgetting. The domain shifts caused by variations in imaging modalities and protocols make robust adaptation even more challenging. We argue that effective cross-domain adaptation requires constraining the model to rely on domain-invariant 3D geometry, rather than domain-specific textures.To address these challenges, we present Kolmogorov-Arnold Shape-aware Segment Anything Model (KS-SAM), a dual-engine framework that enforces geometry-driven adaptation for the Segment Anything Model (SAM). A domain-expert 3D network performs Uncertainty-Aware Feature Fusion (UAFF), adding noise to the input CT image and performing K forward passes to refine the SDF priors. These geometry-aware priors are injected into SAM via the Kolmogorov-Arnold-driven Prior-Semantic Synergistic Distillation (KPSSD) module, enabling bidirectional geometric-semantic anchoring without leaking appearance statistics. In parallel,Kolmogorov-Arnold-driven Low-Rank Adaptation ( KAN-LoRA ) introduces spline-based localized activations to model nonlinear modality shifts while mitigating catastrophic forgetting. Across multi-center and public benchmarks, KS-SAM significantly outperforms state-of-the-art methods on unseen domains, demonstrating that geometry-anchored adaptation is critical for reliable 3D medical SDG.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1258_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{WuRen_KSSAM_MICCAI2026,
author = { Wu, Rengmin AND Li, Jiaqiang AND Luo, Junqin AND Wei, Yuxing AND Liang, Dan AND Zhou, Hui AND Li, Guangming AND Wei, Xinhua AND Wang, Tianfu AND Lei, Baiying},
title = { { KS-SAM: Kolmogorov-Arnold-driven Shape Alignment for Single-Domain Generalization in 3D Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, authors proposed Kolmogorov Arnold Shape-aware Segment Anything Model (KS-SAM), a dual-engine interactive framework tailored for 3D medical Single-Domain Generalization SDG.
- 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.
Their approach is tested on two databases and compared with five SOTA 3D segmentation methods. Their experimental results across several datasets seem reasonable and promising.
- 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.
-SDF is not defined in the abstract. -The authors should study the statistical significance of the classification results of the different approaches, i. e. p-value. -The authors should provide the code of their proposed approach for transparency. -The qualitative comparison is not clear between the different approaches: what is the ROI that was segmented from the CT slice ?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 authors should study the statistical significance of their results and provide the code for transparency.
- 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 KS-SAM, a dual-engine framework for single-domain generalization in 3D medical image segmentation. It has three main components. UAFF uses a 3D UNet with K stochastic forward passes on noise-augmented inputs to produce uncertainty-filtered Signed Distance Field priors. KPSSD injects these SDF priors into SAM via a four-way bi-attention scheme (S→S, S→U, U→S, U→U) followed by a KAN-based layer for non-linear SDF distillation. KAN-LoRA replaces standard linear LoRA projections with a parallel KAN bottleneck using B-spline activations for non-linear adaptation. The framework is evaluated on a multi-center appendix cohort (1,239 cases, four centers) and a public kidney benchmark (WORD→BTCV/KiTS23), reporting substantial Dice gains over nnUNet, nnFormer, UNETR++, VNet, and MedSAM3D.
- 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 core intuition is sound. Anchoring adaptation on SDF-based geometry rather than intensity statistics is a reasonable response to the known SAM failure mode on medical volumes, and the paper executes on this premise end-to-end. The empirical gains are large and not marginal. A mean Dice of 66.56 vs 61.83 on the appendix benchmark and 89.29 vs 84.00 on the kidney benchmark is a meaningful jump. The NSD improvements are even more striking (24.41 → 8.31 mm on appendix), which is consistent with the SDF-driven boundary supervision story. The multi-center appendix dataset with four centers is a genuinely useful SDG testbed. Evaluating on three unseen domains plus a public kidney cross-dataset generalization (WORD→BTCV, WORD→KiTS23) is the right protocol. The ablation cleanly isolates the three components on Domain F and shows KPSSD contributes the largest gain (+3.55 Dice), which matches the paper’s central claim that geometric priors are the key ingredient. The uncertainty-filtered fusion is a sensible way to stabilize SDF priors, and combining segmentation entropy with SDF epistemic variance into adaptive confidence weights is well-motivated.
- 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 most serious issue is missing SDG baselines. The paper compares only against generic 3D segmentors (nnUNet, nnFormer, UNETR++, VNet, MedSAM3D), none of which are SDG methods. There are no comparisons against recent single-domain generalization baselines — no SLAug, no MaxStyle, no CSDG, no teacher-student adversarial augmentation work (which is actually cited as [12] but not compared against). Without at least two dedicated SDG baselines, the “significantly outperforms SOTA” claim on SDG specifically is not supported. The ablation is too narrow. It is conducted only on Domain F (BTCV), not across all OOD targets. The framework has three components but the table is missing key interactions (KAN-LoRA + KPSSD without UAFF, full model without KAN-LoRA), and the ordering of rows makes it hard to separate marginal from cumulative contributions. A full 2³ factorial or at least ablations on a second OOD domain are needed. Several design choices are stated but not justified empirically. The progressive curriculum for K and σ is mentioned but neither the schedule nor its ablation is reported. K itself is never reported as a number — Table 1 does not say whether K=5, 10, or 20.The KAN-LoRA B-spline grid size, spline order, and rank r are all unspecified. These are exactly the hyperparameters that determine whether the method is reproducible. The KPSSD four-way attention formulation in Eq. 2 is written as a single unified matrix operation, but with Q=[Q^S;Q^U] and a shared K/V, this collapses to a standard cross-attention over concatenated queries. The claim that this is a novel “quadruple scheme” modelling intra- and inter-domain correlations is not supported by the equation as written. Either the equation is incomplete or the novelty claim needs to be softened. If authors provide the clarification in rebuttal, I’m happy to increase the score. Good luck.
- 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 motivation is strong and the reported gains are large, especially on boundary metrics. But the empirical case has gaps that matter for an SDG-focused paper. No dedicated SDG baselines are compared against. The ablation is on a single domain. Key hyperparameters are unspecified and the KPSSD equation does not support the novelty claim as written. There is also an apparent arithmetic inconsistency in Table 2.With SDG baselines added, a broader ablation, and the formulation clarified, this moves to accept. As submitted, the claims are broader than the evidence.
- 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 #3
- Please describe the contribution of the paper
The paper addresses the challenging problem of Single-Domain Generalization (SDG) in 3D medical image segmentation. It identifies key limitations in existing approaches, particularly their reliance on texture features and limited ability to capture domain-invariant anatomical structures.
To overcome these issues, the authors propose the Kolmogorov–Arnold Shape-aware Segment Anything Model (KS-SAM), a novel framework that integrates structural priors into a foundation model. The method leverages consensus-aware Signed Distance Fields (SDFs) to enforce anatomical consistency across domains while reducing texture leakage. In addition, the incorporation of KAN-LoRA enables flexible modelling of complex intensity distributions and mitigates catastrophic forgetting, preserving the generalization capabilities of the base model.
The paper further demonstrates the effectiveness of the proposed method on multi-center datasets, showing improved segmentation accuracy (Dice score) and reduced performance variance compared to existing approaches.
- 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 tackles a highly relevant and challenging problem in medical image analysis, particularly SDG, which is underexplored yet critical for real-world deployment.
The proposed KS-SAM framework is conceptually novel, especially in its use of Signed Distance Fields to enforce domain-invariant anatomical structure. This represents a meaningful shift from texture-based generalization approaches.
The integration of geometric priors with SAM is both original and timely, aligning well with current trends in leveraging large pre-trained models.
The use of KAN-LoRA is an interesting and well-motivated design choice, providing both flexibility in modelling intensity distributions and a principled way to mitigate catastrophic forgetting.
The evaluation on multi-center datasets strengthens the practical relevance of the work and demonstrates robustness to real-world variability.
- 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.
While the method emphasizes structural consistency, its performance in cases with significant anatomical abnormalities or rare pathologies is unclear, where shape priors may not generalize well.
The computational cost and scalability of the approach are not discussed, particularly given the additional processing required for SDF extraction and spline-based modelling.
Although the dataset is multi-center and diverse, it remains limited in scale relative to the complexity of the proposed multi-layer architecture. Larger datasets would likely be necessary to fully validate robustness and generalization.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 paper presents a novel and well-motivated approach to Single-Domain Generalization in 3D medical image segmentation, with a clear methodological contribution and strong empirical results. The integration of structural priors via SDFs and the use of KAN-LoRA within a foundation model framework are particularly innovative and relevant.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
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.
This manuscript received three positive reviews, showing the superior novelty, comprehensive evaluation and well-organized structure. Therefore, a provisional accept is given based on the reviews.