List of Papers Browse by Subject Areas Author List
Abstract
Intra-domain distribution shifts are common in medical image segmentation due to variations in acquisition protocols, reconstruction settings, and anatomical morphology, even when data originate from the same nominal domain. We propose a fully automatic segmentation framework that explicitly models latent sub-domains without gradient-based adaptation at inference. Our method dynamically discovers sub-domains via a nonparametric prototype mechanism, expanding only when novel distributional modes are detected. Each prototype conditions a specialized expert via a hypernetwork, forming a prototype-conditioned expert field over the latent manifold. Soft routing with uncertainty-based temperature scaling enables parameter-free test-time adaptation by calibrating sub-domain posteriors without updating model weights. Experiments on 2D medical benchmarks demonstrate improved robustness and granular segmentation accuracy under intra-domain shifts.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6534_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/lsd-seg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LeeKyu_Dynamic_MICCAI2026,
author = { Lee, Kyungsu AND Choi, Seo-Yeon AND Hwang, Jae Youn AND Woo, Jonghye},
title = { { Dynamic Sub-domain Modeling for Robust 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
The paper proposes a sub domain modeling framework for robust medical image segmentation under intra domain distributional heterogeneity. The key contributions are a formal problem reformulation that treats single domain segmentation as inference over a latent sub domain mixture variable. Merging a hypernetwork that maps each prototype to lightweight adapter deltas for a shared SAM-style decoder. Uses an implicit expert field over the sub domain manifold an uncertainty tempered soft routing that modulates expert mixing through disagreement based temperature scaling enabling fully parameter free test time inference. The framework is evaluated across 20 public segmentation benchmarks spanning CT, MRI, endoscopy, dermoscopy, ultrasound, fundus, histopathology and PET and CT outperforming strong baselines including nnU-Net, MedNeXt, Mamba-based models.
- 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.
Strengths – Framing intra domain heterogeneity as a latent mixture is underexplored in medical image segmentation. – The ability to determine K from data with bounded growth, merging, and pruning is a good technical advance over standard MoE approaches. It avoids the need for prior knowledge about the number of sub domains. – The strict constraint of no test-time gradient updates is clinically important for deployment. – 20 datasets across 6+ imaging modalities under a unified 2D protocol, with worst case sub domain Dice and Boundary F1 alongside standard metrics is comprehensive – All four main components dynamic vs. fixed-K discovery, hypernetwork vs. static adapters, uncertainty tempered vs standard routing, ROI vs global only descriptor are individually ablated and shown to contribute to the design choices. – The UMAP visualization and uncertainty and routing analysis in Fig 2 and 3 provide genuine insight into what the model has learned going.
- 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.
Weaknesses – Fig. 2(b) appears uniformly positive across all 20 datasets and all 13 baselines, with no losses shown. The authors must provide per dataset breakdowns and appropriate statistical significance testing for example Wilcoxon signed rank. – For datasets lacking metadata, sub domains are constructed by clustering training image descriptors. Since the same descriptor network contributes to both sub domain definition and model behavior, this risks circular evaluation of robustness. – The paper explicitly claims to report runtime and peak GPU memory, but these figures are absent from all tables and figures. – UMAP visualization is underspecified. Fig. 2(a) shows only Dataset A and Dataset B without identifying which of the 20 datasets these correspond to. – The paper does not discuss scenarios where the framework may underperform
- 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.
(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 addresses an important problem of intra-domain heterogeneity as a structural ceiling in medical image segmentation with a technically sound and novel framework. The evaluation scope across 20 datasets is impressive, and the focus on worst case sub domain Dice is clinically well motivated. However, several concerns prevent a stronger recommendation, please refer to weaknesses section.
- 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
Review #2
- Please describe the contribution of the paper
1.This paper addresses an important and underexplored problem, which is intra-domain heterogeneity in medical image segmentation. 2.Evaluation results on 20 datasets surpass other competitors.
- 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.Well-motivated problem framing. The paper points out a common but under-addressed failure mode, even within one dataset, shifts from scanner/protocol/patient morphology can create a mixture of sub-distributions that a single model averages over. 2.The proposed architecture is technically sound. For example, modulating the routing temperature based on expert disagreement is a clever mechanism to prevent overconfident predictions on ambiguous boundary cases. 3.The evaluation includes 20 distinct public datasets spanning multiple modalities (MRI, CT, Ultrasound, Endoscopy, Dermoscopy, etc.). providing evidence of the method’s generalizability.
- 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.In Section 4 (Settings), the authors state: “runtime and peak GPU memory are reported.” However, this data is completely missing from the provided text. 2.Since the network relies on automatic prompts to generate local descriptors, the paper lacks details on how robust the system is to poor-quality automatic prompts. If the initial prompt generation fails or is highly inaccurate, does the sub-domain descriptor “r” collapse, leading to incorrect expert routing? An ablation or discussion on prompt sensitivity would strengthen the paper. 3.Considering the substantial heterogeneity across the 20 datasets, particularly in modality and degree of intra-domain variation, it is important to clarify whether the key hyperparameters τ, ρ, and α were kept fixed across all datasets or tuned separately for each one. If a single hyperparameter setting works well across 20 highly diverse datasets, that would be a major strength and should be explicitly highlighted. If, instead, these parameters were tuned per dataset, the paper should clearly describe the tuning protocol and discuss the associated overhead when transferring the method to a new dataset, especially relative to methods such as SAM2/MedSAM2 that do not appear to require this level of dataset-specific tuning. 4.The notation around the temperature scaling is currently under-specified. the paper does not clearly define either T_0 or the subscript r in T_r. 5.A key concern is how the method behaves when a genuinely new sub-domain appears at test time. Because the prototypes are fixed during inference, the model can only reweight known experts and cannot explicitly instantiate a new prototype for an unseen mode. 6.Figure 2(a) does not convincingly support the paper’s core claim of meaningful sub-domain discovery. While Dataset A shows some clustered structure, Dataset B appears largely as a homogeneous blob with no clear sub-domain separation. Thus, it remains unclear whether the reported gains come from explicit sub-domain modeling or simply from the increased capacity introduced by the hypernetwork, multiple experts, and routing mechanism. The current results and ablations do not sufficiently disentangle semantic expert specialization from a generic MoE/ensemble effect.
- 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.
(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?
Although this paper is well-motivated and focuses on an interesting and practical problem in medical image analysis. The key concern is that the paper does not convincingly show that the performance gains come specifically from latent sub-domain discovery, rather than simply from the increased capacity and flexibility of the prototype-conditioned multi-expert architecture. The evidence for meaningful sub-domain structure is indirect, with no sufficiently direct visualization or analysis of learned sub-domains, routing behavior, or expert specialization. In addition, the paper lacks important implementation details needed for reproducibility and fair assessment.
- 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 paper proposes a Dynamic Sub-domain Modeling framework that treats single-dataset medical image segmentation as inference over latent sub-domains, using a frozen SAM encoder, task-specific descriptors, and nonparametric dynamic prototypes to handle intra-dataset heterogeneity. A hypernetwork-generated expert field with an uncertainty tempered soft routing mechanism enables gradient free inference, and the method achieves strong performance across 20 public datasets, with average Dice improvements of about +3.9 over strong baselines and overall Dice 87.9 / HD95 11.8.
- 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.
- Conceptual novelty and perspective: The work reframes intra-domain variability (scanner, protocol, anatomy differences) as meaningful structure rather than noise, dynamically assigning specialized experts to discovered sub-domains and improving boundary delineation (e.g., Boundary F1).
- Dynamic, nonparametric sub-domain discovery: Nonparametric prototypes allow the number of sub-domains to grow as new modes appear during training, avoiding the need to predefine sub-domain counts and enhancing adaptability.
- Inference-time practicality: The framework is fully gradient free at test time, relying on the learned prototypes and routing mechanism without parameter updates, which is attractive for stable and efficient clinical deployment.
- Broad and convincing validation: The method is evaluated on 20 datasets across multiple imaging modalities and consistently outperforms strong baselines such as nnU Net 2D, SegMamba, and MedSAM2+APG, with supporting analyses (e.g., UMAP) to illustrate generalization.
- Robust performance under domain shifts: The dynamic combination of experts yields high Dice (87.9) and competitive HD95 (11.8) across heterogeneous domains, including previously unseen distributions.
- 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.
- Limited architectural/detail transparency: Key components such as the prototype memory bank, the structure of expert models, and the mechanism by which experts expand remain under-specified, which hinders reproducibility and makes it hard to reimplement the method faithfully.
- Sensitivity to thresholding and noise: Prototype creation hinges on a threshold parameter τ; despite pruning/merging, real world noisy clinical data may require careful tuning, weakening claims of full automation.
- Dependence on SAM-based descriptor space: Sub-domain descriptors are derived from a frozen SAM encoder, so if this backbone fails to capture highly specialized medical textures, sub-domain discovery and expert specialization may be suboptimal.
- Lack of explicit 3D modeling: The approach operates slice wise in 2D and does not explicitly enforce inter-slice or volumetric consistency, which is important for many 3D medical segmentation tasks.
- Incomplete statistical analysis: While mean values and 95% confidence intervals are reported, there are no formal significance tests (e.g., paired t tests or Wilcoxon tests), so the strength of claims over baselines is not rigorously quantified.
- 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?
I recommend a weak accept because the paper offers a fresh and practically meaningful approach to robust medical image segmentation under intra-dataset heterogeneity, with clear empirical gains, but also has notable gaps in technical detail and analysis. The idea of dynamic sub-domain modeling with nonparametric prototypes, coupled with a hypernetwork-generated expert field and uncertainty tempered routing, is conceptually strong, leads to gradient free inference, and yields consistent improvements over well-established baselines on 20 diverse datasets. At the same time, the reproducibility concerns (insufficient description of the prototype memory and expert expansion), sensitivity to the threshold τ, lack of explicit 3D consistency, and missing statistical significance tests prevent a full, unreserved acceptance. Overall, the contribution is novel and promising, and the experimental results are compelling enough that, despite these limitations, the work is suitable for publication with revisions that clarify the architecture, better analyze prototype behavior and scaling, and strengthen statistical validation.
- 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
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 work received overall positive comments (with two reviewers recommending weak accept and one reviewer suggesting weak reject). Reviewers acknowledge the novelty of the proposed framework, and the comprehensive evaluation on 20 public segmentation benchmarks, as well as the importance/relevance of the addressed problem. On the other hand, the lack of important details for its reproducibility, sensitivity analysis to several hyperparameters, and the absence of evidence that the reported gains genuinely stem from latent sub‑domain discovery undermine its overall impact. Given the limited space in the MICCAI format, and the overall positive comments, I recommend the acceptance of this work. Nevertheless, I strongly suggest the authors to address these comments in the revised version (or even in a GitHub public repository for the reproducibility concerns).