List of Papers Browse by Subject Areas Author List
Abstract
Ambiguous medical image segmentation aims to provide a series of diverse but plausible segmentation hypotheses. However, existing methods introduce stochasticity in a fixed and pre-defined manner, failing to form a progressive semantic modeling process. To address these challenges, we propose KANResDiff to learn local residual diffusion with Kolmogorov-Arnold Network, thereby assigning distinct roles across stages for ambiguity modeling. Specifically, we propose Independent Time Encoding that offers spline-based time embeddings instead of linear ones from MLPs, which enhances the independence across inference stages and assigns progressive semantic roles to different stages. We propose Residual Schr"odinger Bridge that injects deterministic residual prior with learnable weights by constructing local Schr"odinger Bridge instead of following manually settings, achieving a flexible deterministic–stochastic interaction and stage-aware ambiguity modeling thanks to local optimal diffusion path. Extensive experimental results on two public datasets demonstrate that KANResDiff achieves SOTA performance on GED and HM-IoU, with maximum improvements of $16.8\%$ and $7.7\%$, respectively, while maintaining competitive performance on the MDM metric. Source code is available at \url{https://github.com/PerceptionComputingLab/KANResDiff}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0621_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/PerceptionComputingLab/KANResDiff
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiFan_KANResDiff_MICCAI2026,
author = { Li, Fanding AND Wang, Chenglin AND Li, Xiangyu AND Qiu, Xingyu AND Ma, Xinghua AND Yin, Xiangming AND Li, Haiyang AND Dong, Suyu AND Wang, Wei AND Wang, Kuanquan AND Luo, Gongning AND Li, Shuo},
title = { { KANResDiff: Learning Local Residual Diffusion via Kolmogorov-Arnold Network for Ambiguous 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 paper proposes a KANResDiff including an independent time encoding and a residual schrodinger bridge for AMIS. A ITE is introduced to enhance the independence across different timesteps and facilitates a progressive ambiguity modeling process. Besides, the proposed RSB is proposed to enabling flexible and adaptive determinism injection into the diffusion process. Experiments on two medical image segmentation datasets are conducted to verify the effectiveness of the proposed method.
- 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 strengths of this paper include: First, the KANResDiff proposed in this paper is highly targeted to address specific problems in ambiguous medical image segmentation. Besides, comprehensive motivations are introduced and promising experimental results are produced by the proposed method.
- 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、Although the method introduces multiple components with interpretability motivations (e. g. , ITE and RSB), the overall framework becomes highly complex, making the final decision process still difficult to interpret in a clinically meaningful way. 2、The combination of diffusion modeling, KAN-based function representation, and step-wise residual estimation significantly increases computational overhead, which makes theproposed method less suitable for real-time or resource-constrained clinical applications.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 strong motivation by associating diffusion timesteps with different annotations. However, the overall design is relatively complex. By the way, by introducing the extra stochastic inferences, how to ensure the convergence of the diffusion model?
- 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 you for the author’s response. For ambiguous image segmentation, linking the time steps with different segmentation granularities while retaining a fixed branch is reasonable. Most of my concerns have been addressed, so I will maintain my original score.
Review #2
- Please describe the contribution of the paper
This paper proposes KANResDiff, a diffusion-based framework for ambiguous medical image segmentation. The method introduces two components: Independent Time Encoding (ITE), which replaces MLP-based time embeddings with B-spline parameterizations via Kolmogorov-Arnold Networks (KANs) to enhance cross-timestep independence, and Residual Schrödinger Bridge (RSB).
- 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 paper identifies a concrete limitation in existing residual diffusion methods: shared MLP time embeddings create cross-timestep gradient coupling that blurs the semantic roles of different diffusion stages. This observation is formalized through a gradient analysis in Eq.(1–5) rather than left as a narrative claim, which provides a mathematically substantiated foundation for the proposed ITE module. 2.By exploiting the local support property of B-spline basis functions, ITE ensures that time embedding updates at timestep t_i do not propagate to the representation of distant timesteps. 3.RSB replaces heuristic residual weighting with a principled optimal transport formulation. It provides a data-driven, geometry-aware mechanism for balancing stochasticity and deterministic structure across diffusion timesteps.
- 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 central technical idea of combining KANs with local Schrödinger Bridges is already proposed in Qiu et al., “Finding Local Diffusion Schrödinger Bridge using Kolmogorov-Arnold Network,” CVPR 2025 (Reference [15] in the submission). RSB appears to extend [15] by adding a residual deterministic prior with learnable weights, but this distinction is never explicitly analyzed. Without a rigorous technical comparison against [15], the contribution risks being perceived as a task-specific adaptation of an existing method rather than an independent advance. 2.The application of Schrödinger Bridge to AMIS is not novel. The paper implicitly positions itself as pioneering the use of Schrödinger Bridges for AMIS, yet SSB (Baru et al., MICCAI 2025, Reference [1]) already establishes this connection. The incremental novelty over SSB reduces to KAN-based time encoding and dynamic residual weighting. 3.The rationale for adopting KAN is insufficiently justified. The paper never adequately answers why KAN is specifically necessary to address this need. B-spline basis functions and is not unique to the KAN formulation. Alternative bases with equivalent or stronger local support, such as Radial Basis Functions (RBFs) or locally-windowed Fourier features, could plausibly achieve the same decoupling effect. The authors provide no ablation study or theoretical argument demonstrating that KAN is superior to these alternatives in the AMIS context. As a result, the adoption of KAN reads as an engineering choice motivated by novelty rather than a principled design decision. 4.MDM performance regression is not adequately addressed.In Table 1, KANResDiff achieves MDM = 0.930, underperforming SSB (0.942) and P2SAM (0.939). The authors dismiss this as “competitive performance,” which is misleading.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
1.The central technical idea underlying both ITE and RSB — combining KAN’s B-spline parameterization with local Schrödinger Bridges in a diffusion framework — is already present in that concurrent work. 2.The unjustified selection of KAN over functionally equivalent alternatives. 3.The underperformance on MDM is neither explained nor discussed.
- 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 author’s feedback solved my concerns.
Review #3
- Please describe the contribution of the paper
In this work, the authors aim to address existing limitations of ambiguous segmentation methods, namely that existing methods introduce stochasticity in a fixed/pre-defined manner, which limits the representation of heterogeneous ambiguity. Addressing this, the authors propose KANResDiff, which incorporates Kolmogorov-Arnold Networks (KANs) with residual diffusion for ambiguity modeling. Their method specifically utilizes their proposed Independent Time Encoding (ITE), which replaces the linear embeddings produced by multilayer perceptrons (MLPs) with spline-based time embeddings. This helps increase independence across inference stages. The authors also propose a Residual Schrödinger Bridge that injects deterministic residual prior with learnable weights into the network.
- 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.
To me, this is a very good paper. The authors do a good job of introducing the problem(s) that they are trying to solve (namely, the lack of independence between inference states and limited ambiguity-awareness during inference), and the methods that they propose seem genuinely novel in this domain. I have not seen spline-based KANs used in this manner before for ambiguous segmentation, and the integration of the Residual Schrödinger Bridge further enhances the novelty. The results are convincing in terms of the performance improvement over the baselines offered by KANResDiff. On top of this, the paper is very easy to read, and the figures, particular Figs 1 and 2, are high-quality and enhance the interpretability of the paper.
- 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.KANs can be implemented in various different ways, particularly in terms of their basis functions. In this work the authors utilize B-splines, which I imagine indicates that they use the KAN implementation used by Liu et al. [1] when they originally proposed KANs. However, several KAN variants using different basis fucntions have gained popularity, including those utilizing radial basis functions [2], wavelet transformations [3], and Chebyshev polynomials [4], just to name a few. At the moment, I feel the paper is lacking in a discussion on whether the authors’ method is reliant solely on B-splines as the basis function, or whether they think it would be generalizable to other KAN variants with different basis functions. In the event that it is the splines themselves, not KANs, that make the method work, it would be potent to discuss whether one could simply use a spline-based activation function in an MLP, which has been a topic of great debate in discussions regarding KANs (see [5], for example).
2.While I praised the quality of Figs. 1 & 2 in the strengths, I think what Fig. 3 depicts is a little unclear at the moment. Obviously, it’s a qualitative comparison between KANResDiff, the ground truth, and two baseline models, but I’m not entirely sure I understand what the red boxes around certain masks are. I assume the intent is to demonstrate that KANResDiff produces visually better masks compared to the baselines and/or better alignment to the ground truth distribution, but I can’t quite figure out how the red boxes are depicting that currently. To me, they seem drawn randomly.
3.Staying on the topic of visualization, the authors provide visualizations for how KANResDiff performs on the LIDC-IDRI dataset, but not for the ISIC3 dataset. It would have been nice to have visualizations on both datasets.
References: [1] Liu, Ziming, et al. “Kan: Kolmogorov-arnold networks.” arXiv preprint arXiv:2404.19756 (2024). [2] Ta, Hoang-Thang. “BSRBF-KAN: A combination of B-splines and radial basis functions in Kolmogorov-Arnold networks.” International Symposium on Information and Communication Technology. Singapore: Springer Nature Singapore, 2024.[3] Bozorgasl, Zavareh, and Hao Chen. “Wav-kan: Wavelet kolmogorov-arnold networks, 2024.” arXiv preprint arXiv:2405.12832.[4] SS, Sidharth, Keerthana AR, and Anas KP. “Chebyshev polynomial-based kolmogorov-arnold networks: An efficient architecture for nonlinear function approximation.” arXiv preprint arXiv:2405.07200 (2024). [5] Dhiman, Vikas. “Kan: Kolmogorov–arnold networks: A review.” Retrieved from (2024).
- 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.
(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?
I think this work is very nicely presented, and proposes novel contributions in the area of ambiguous medical image segmentation. I have not seen any work utilizing KANs for ambiguous medical image segmentation, especially not in the way the authors have used them with the independent time encoding. Likewise, the residual Schrodinger’s bridge appears novel, and seems to help the performance of the model.
- 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.
I thank the authors for their diligent rebuttal. I find their response to my comments satisfactorily, as well as their response to the other reviewers. Particularly, I think they sufficiently address the concerns raised by Reviewer 2 on the point of novelty in a way that does not hamper the claims of their own work. Because of this, I choose to keep my pre-rebuttal rating of Accept. I would, however, encourage the authors, if the work is accepted, to include in their camera ready version a discussion on how an MLP trained with a B-spline activation function would perform on the task explored by the authors in place of KANs. I think such a discussion is necessary for this work, and would be of great value to the paper and of interest to readers.
Author Feedback
Thank all reviewers for the positive comments including ‘a very good paper’ ‘genuinely novel’ ‘results are convincing’ ‘easy to read’ ‘high-quality figures’ ‘comprehensive motivations’ ‘promising experimental results’ ‘identifies a concrete limitation’ ‘provides a data-driven, geometry-aware mechanism’.
Reviewer 1: 1.In our KANResDiff, ITE and RSB are jointly designed as complementary components for progressive ambiguity modeling rather than independent modules. KANResDiff performs anatomical consistency modeling in early stages and ambiguity diversification in later stages, aligning with clinicians’ coarse-to-fine ambiguity refinement process and demonstrating clinical interpretability. 2.KANResDiff requires 0.36s to generate one prediction with a model size of 670MB, compared with 11s and 485MB for CCDM [19], demonstrating real-time inference efficiency with moderate memory overhead. Specifically, ITE replaces the original MLP time embedding with comparable scale while RSB requires only additional training procedures without increasing inference complexity. 3.The convergence of KANResDiff is ensured by constraining optimization around a pretrained and converged diffusion model. Specifically, RSB refines diffusion paths through dynamic residual corrections, which is bounded by the original pretrained diffusion process.
Reviewer 2: 1.Our KANResDiff independently advances [15] by injecting dynamic segmentation-prior residuals into diffusion formulation rather than solely optimizing noise schedules (αt, βt), enabling anatomically consistent refinement and progressive ambiguity evolution for AMIS. Novelty of KANResDiff is that KANResDiff constructs stage-wise local Schrödinger Bridges with dynamic deterministic residual priors, enabling adaptive determinism injection into diffusion beyond the capability of [15]. 2.We respectfully clarify that the claimed novelty of KANResDiff lies in pioneering a new Schrödinger Bridge (SB)-based paradigm for progressive semantic ambiguity evolution rather than merely introducing SB into AMIS. KANResDiff advances [1] through dynamic local SB-guided residual diffusion path refinement, enabling flexible deterministic–stochastic coordination for ambiguity modeling in AMIS. 3.Motivation of KAN: KANs with localized basis functions (e.g. B-splines, RBFs and locally-windowed Fourier features) consistently exhibit continuous temporal evolution and reduced forgetting effects. KANs provide a principled implementation for progressive ambiguity modeling in AMIS through localized timestep parameterization, including independent time-embeddings optimization and local residual weights assigning. Exploring alternative KAN variants will be our future work. 4.MDM evaluates individual accuracy, while GED and HM-IoU evaluate annotation distribution fitting. KANResDiff establishes a principled semantic ambiguity modeling trajectory, achieving superior GED and HM-IoU while maintaining competitive MDM with marginal differences from SSB and P2SAM.
Reviewer 4: 1.Motivation of KAN: KANs with localized basis functions [1,2,3,4] consistently exhibit continuous temporal evolution and reduced forgetting effects compared with globally shared MLPs. Based on these properties, our KANResDiff pioneers a progressive ambiguity modeling paradigm for AMIS and establishes an extensible framework for future KAN variants with different localized basis functions. 2.Within the highlighted regions in Fig. 3, KANResDiff generates anatomically and semantically plausible structures following ground-truth patterns, while baseline methods may exhibit inconsistent boundaries or anatomically unreasonable shapes. 3.Qualitative evaluations on both datasets demonstrate KANResDiff achieves better alignment with ground-truth patterns and higher clinical plausibility. We mainly present LIDC visualizations that more clearly highlight the effectiveness of KANResDiff, while ISIC3 visualizations will be included in the final released code repository.
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 are split. Two are positive about the mathematically grounded gradient-coupling motivation, the local-support property of B-spline time encoding, and the headline gains in GED and HM-IoU. One is moderately negative on grounds of incremental novelty and incomplete prior-art positioning, raising a substantive concern that merits examination through rebuttal: the technical core — KAN-parameterized local Schrödinger Bridge for diffusion — is already published in Qiu et al. (CVPR 2025), and Schrödinger-Bridge-for-AMIS is in turn established in Baru et al. (MICCAI 2025/SSB).
For the rebuttal, authors should engage directly with the prior-art and ablation points raised by the negative reviewer — most importantly, the relationship to Qiu et al. (CVPR 2025) and SSB (Baru MICCAI 2025), and the basis-function alternatives noted by both the negative and one positive reviewer. Further items that would strengthen the response include a head-to-head comparison with SSB across all metrics, addressing computational overhead, and proper attribution of the MLP time-embedding pattern.
- 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 proposes a KAN-parameterized local Schrödinger Bridge for ambiguous medical image segmentation, together with spline-based time encodings from KAN that enhances independences across inference, with headline gains in GED and HM-IoU supporting the approach.
The reviewers were initially split, with the principal concern centering on the differentiation of the technical core from Qiu et al. (CVPR 2025) and the Schrödinger Bridge formulation in Baru et al. (MICCAI 2025/SSB). The authors’ rebuttal addressed these points to the reviewers’ satisfaction, and all reviewers converged on a consensus to accept. I concur with this consensus in recommending 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.
Reviewers summarized strengths of your work, novel aspects and noted that your solution to the specific application area may be of potential interest. The rebuttal helped to clarify open issues and misunderstandings. Would the paper be accepted, authors are strongly encouraged to improve the final write-up by including recommendations by reviewers, in particular improving clarity of notations and descriptions.
Meta-review #3
- 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.
All reviewers recommended a clear acceptance of this work after the rebuttal. The main contribution of this work is the claimed “progressive semantic ambiguity evolution” based on the Schrödinger Bridge. The authors attempt to leverage more modern techniques to improve ambiguous medical image segmentation, which could be beneficial to the community. Therefore, a final acceptance is recommended, and the authors are suggested to address the reviewers’ comments in the camera-ready version.
