List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised learning (SSL) has become a pivotal strategy for medical image segmentation by harnessing unlabeled data. However, most consistency-based paradigms focus on aligning pixel-level outputs, failing to capture the high-order structural knowledge and diagnostic logic essential for complex anatomical interpretation. Motivated by the hierarchical reasoning process in clinical diagnosis—where radiologists progressively integrate local cues, global context, and anatomical priors—we propose \textbf{SKTC-Net}, a novel framework centered on \textbf{Structured Knowledge Topology Consistency}. We reformulate the semi-supervised objective by mimicking three critical aspects of clinical diagnosis: (i) \textbf{Where to look}: \textit{Spatial Attention Consistency} aligns saliency maps to ensure consensus on salient anatomical regions; (ii) \textbf{How to compare}: \textit{Relational Structure Consistency} preserves the global topological distribution among learning samples to mimic comparative diagnosis; and (iii) \textbf{How to reason}: \textit{Evolutionary Flow Consistency} constrains the inter-layer feature transitions to maintain a stable diagnostic reasoning path. Evaluations on ACDC (2D) and Pancreas (3D) datasets demonstrate that SKTC-Net consistently performs superior. Notably, SKTC-Net achieves consistent gains in low-label settings (e.g., +3.47\% Dice at 10\% Pancreas), attributed to the complementary supervision from multi-level structural constraints.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3226_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SSonnyboy/STKC-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheYu_Learning_MICCAI2026,
author = { Chen, Yu AND Xu, Qiang AND Wang, Ke AND Yang, Fan AND Wei, Honghao},
title = { { Learning Beyond Pixels: Structured Knowledge Topology Consistency for Semi-supervised 2D/3D Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a new semi-supervised medical image segmentation method. The authors argue that existing methods mostly focus on output pixel level consistency while ignoring high order structural knowledge. Their method enhances semi-supervised learning by simulating the clinical reasoning process.
- 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 three proposed modules are basically well designed with implementation details and applicability.
The experiments are comprehensive and the performance improvements are notable.
Existing models can directly adopt the proposed modules.
The paper is well organized and clearly written.
- 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 innovation of core components such as SAC may not be as significant as the authors claim. Similar ideas have been widely discussed in the community, for example Attention Transfer (ICLR 2017).
2.The experiments only use MRI data. The performance on other modalities such as ultrasound remains questionable.
3.The paper claims applicability to both 2D and 3D tasks, but the method is introduced from a 2D perspective without clear explanation of how it adapts to 3D data.
4.Some theoretical connections are not clearly explained. For instance, RSC enforces consistency on the inter sample similarity in feature space, but the authors relate it to anatomical structure preservation without sufficient justification.
- 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?
This paper is well written, and i find it satisfying to read. However, i have some minor concerns.
1.RSC constrains the similarity between samples in a batch rather than direct anatomical structures. Why can the former be equated to the latter? Does RSC incorporate spatial location information of organs, and is there functional overlap with SAC?
2.The ablation study does not isolate the individual contribution of each loss component. Only sequential addition is shown.
3.The three constraints may contribute differently across different organs and modalities. Could an adaptive weighting mechanism be designed instead of a fixed α?
4.What is the computational overhead of the dual student architecture and the matrix computations? How do the parameter count and inference time compare with existing methods?
5.In Figure 3, the right side only shows conceptual descriptions such as “where to look”. It would be helpful to also indicate how each constraint is implemented, for example “agreement between spatial saliency maps”.
6.In Section 2, it is unclear whether F_i^l refers to a single sample or the entire batch.
7.Why are MT and CPS used as baselines in Table 5 but not included in Tables 1 and 2?
8.Minor formatting issue: In Equation (1), the use of \mathcal{D} should be consistent.
- 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.
In the rebuttal stage, the authors basically explained/solved my & R2, R3’s concerns. Considering the novelty, although questioned by all of us, the authors’ explanation is acceptable to me. The authors also clarified the potential extension to 3D data and the anatomical prior modeling. So, I’ll keep the ‘accept’.
Review #2
- Please describe the contribution of the paper
This paper proposes SKTC-Net, a semi-supervised segmentation framework that extends consistency regularization beyond pixel-level outputs to structured representations, including spatial attention, inter-sample relationships, and inter-layer feature transitions.
- 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 work is well motivated and extends conventional consistency learning to structured feature spaces through attention, relational, and hierarchical constraints. The method demonstrates consistent improvements across multiple benchmarks and shows good compatibility with existing semi-supervised frameworks.
- 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 is limited and largely incremental, as the method mainly extends standard consistency regularization (e.g., Mean Teacher, CPS) to different feature representations rather than introducing fundamentally new learning principles.
2) The “clinical reasoning” motivation appears largely conceptual, and the proposed modules (SAC, RSC, EFC) can be interpreted as standard attention alignment, feature similarity matching, and feature correlation constraints, which are common in representation learning.
3) The contribution of each component is relatively modest. As shown in the ablation study, individual modules provide only incremental improvements, suggesting limited standalone impact.
4) The method introduces multiple consistency constraints, increasing model complexity without sufficient analysis of computational overhead or training stability. The evaluation is limited to standard benchmarks, without further validation such as cross-dataset generalization or robustness analysis.
- 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?
The paper presents a well-engineered and empirically effective extension of consistency-based semi-supervised segmentation, with consistent improvements across benchmarks. However, the level of novelty is moderate and mainly lies in combining existing ideas rather than introducing fundamentally new concepts, and the experimental validation could be further strengthened.(see weakness)
- 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 rebuttal addresses several of my concerns, especially regarding computational overhead and 3D applicability. While I still consider the novelty somewhat incremental, the method is technically sound, empirically effective, and consistently improves performance across benchmarks. Overall, I believe the paper meets the acceptance bar for MICCAI.
Review #3
- Please describe the contribution of the paper
Motivated by the hierarchical reasoning process of radiologists in clinical diagnosis, this paper proposes a structured knowledge topology consistency network (SKTC-Net), a novel framework for semi-supervised medical image segmentation. Unlike existing methods that rely on redundant network branches, SKTC-Net systematically mimics the clinical diagnostic process across three orthogonal dimensions: (1) Where to look: Spatial attention consistency (SAC) aligns saliency maps to enforce consensus on critical anatomical regions. (2) How to compare: Relation structure consistency (RSC) preserves global topological relationships among samples in a batch by aligning their similarity matrices, mimicking comparative diagnostic logic. (3) How to reason: Evolutionary flow consistency (EFC) constrains inter-layer feature transitions to capture the flow of feature abstraction and maintain a stable reasoning path. The proposed method demonstrates superior performance on both the ACDC and Pancreas 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.
Novel and Well-Motivated Concept: The idea of systematically mimicking the hierarchical clinical reasoning of human experts (local cues, global context, and anatomical priors) without adding redundant network branches is highly novel and logically sound.
Purpose-Driven Architecture: The architectural design is highly cohesive. Each proposed module (SAC, RSC, EFC) aligns perfectly with its intended diagnostic dimension, making the framework intuitively understandable and theoretically robust.
Strong Empirical Performance: The experimental validation is solid. Achieving superior results on recognized benchmarks like ACDC and Pancreas clearly demonstrates the practical effectiveness of the proposed structural consistency constraints.
- 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.
Poor Readability (Sentence Length): The manuscript suffers from overly long and complex sentences, which significantly hinders readability and makes the core concepts difficult to grasp. The manuscript would greatly benefit from being rewritten with shorter, more concise sentences to improve clarity.
Notation Ambiguity: The variable l is used interchangeably throughout the mathematical formulations to denote both “layer” and “labeled” data. This overlapping notation is highly confusing and poses a significant risk of misunderstanding the equations.
Missing Definitions: The loss term L_con is introduced in the text and equations, but is not explicitly defined, leaving a gap in the methodological explanation.
Insufficient Justification for SAC: While SAC effectively focuses on spatial information, the paper lacks a clear comparative analysis with prior spatial attention methods, such as Park et al. [1]. The authors need to explicitly state the novel aspects and structural advantages of SAC over this existing approach.
Lack of Rationale for Cosine Similarity: In the RSC module, cosine similarity is utilized to align similarity matrices. However, it is not clear why this specific metric is chosen over other distance metrics (e.g., L_1 or L_2 norms) or deep learning-based semantic comparison methods (e.g., attention mechanisms or SAM). The authors should elaborate on the specific advantages of using cosine similarity for topological distribution comparison.
Missing Recent Related Works: Several highly relevant and recent papers in the field of semi-supervised medical image segmentation are missing. Including discussions on the following works would strengthen the literature review: (1) Zhu et al., “AdvMIM: Adversarial Masked Image Modeling for Semi-Supervised Medical Image Segmentation”, MICCAI, 2025.(2) Wang et al., “GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation”, ICML, 2025.(3) Liu et al., “Translation Consistent Semi-supervised Segmentation for 3D Medical Images”, TMI, 2025.(4) Peng et al., “DMformer: Difficulty-adapted Masked Transformer for Semi-Supervised Medical Image Segmentation”, BHI, 2025.
[1] Park et al., “Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair”, CVPR, 2024.
- Please rate the clarity and organization of this paper
Poor
- 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?
This paper presents a highly original and thoughtfully designed framework that successfully translates clinical diagnostic logic into a functional semi-supervised learning architecture. The experimental results are robust and clearly validate the superiority of the proposed method. However, the overall readability is heavily compromised by excessively long sentences, and the ambiguous mathematical notations (e.g., overlapping l) create unnecessary confusion. Given the strong core methodology and empirical success, I recommend a Weak Accept. If the authors revise the manuscript to significantly improve clarity, rectify the notation issues, provide the missing justifications (for SAC and cosine similarity), and update the related works.
- 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 authors have adequately clarified the distinction between SAC and Park et al., demonstrating that the two methods differ in both purpose and technical formulation. The justification for cosine similarity in RSC, particularly its scale-invariance and suitability for topological alignment, is reasonable and accepted. Regarding readability, notation ambiguity, and missing definitions, the authors have acknowledged all issues and committed to revising accordingly. I trust that these presentation improvements will be properly reflected in the final version. I encourage the authors to ensure that all promised revisions, including notation fixes, explicit definition of L_con, and incorporation of recent related works, are fully addressed in the camera-ready submission.
Author Feedback
We appreciate the reviewers’ feedback. Following rebuttal guidelines, we’re unable to add new experiments in the responses.
Novelty and Relation to Prior Work (R1, R2, R3): The core contribution is the systematic unification of three orthogonal diagnostic dimensions into a coherent topology consistency framework. SAC differs from Attention Transfer (ICLR 2017): we operate in a dual-student semi-supervised paradigm for unlabeled data consensus, not supervised model compression. The three modules are mutually reinforcing—SAC enforces intra-sample spatial consensus, RSC preserves inter-sample relational topology, and EFC stabilizes inter-layer reasoning dynamics. Table 3 shows synergistic gains, exceeding the sum of individual contributions. Concerning R3’s fundamental new learning principle, MT and CPS frameworks are both introduced into the field of medical image segmentation in semi-supervised classification. Most of the existing excellent work is based on the innovation and modification of the two frameworks, including UAMT, BCP, AD-MT, and so on.
RSC and Anatomical Structure (R1): RSC aligns batch-wise similarity matrices to preserve the global topological distribution of cases in feature space. Anatomically similar patients naturally cluster closer in latent representation; stabilizing this topology mimics radiologists’ comparative diagnostic reasoning. RSC and SAC are functionally orthogonal: SAC operates on spatial attention maps for intra-image localization, while RSC uses flattened global vectors where spatial information is collapsed, preserving only inter-sample relationships.
3D Adaptability and Dataset Modality (R1): All modules naturally extend to 3D. SAC aggregates channels over 3D spatial dimensions; RSC flattens 4D feature tensors; EFC computes Gram matrices between adjacent 3D layers after adaptive pooling. The Pancreas experiments in Table 2 use VNet on 3D CT volumes, explicitly validating 3D applicability. Benchmarks include ACDC (MRI) and Pancreas (CT).
Ablation Study (R1): Sequential ablation demonstrates progressive complementary effects, as modules interact synergistically. Stepwise gains: CPS baseline (80.04 Dice), plus SAC (+1.69), plus RSC (+0.29), plus EFC (+2.20), reaching 84.22.Due to page limits, isolated ablations will be considered in future extended versions.
Computational Overhead (R1, R2): SKTC adds no parameters or branches. Inference uses a single student at the same speed as standard backbone. Training time on ACDC (5% label, RTX 3090): SS-Net 2.44hr, UA-MT 2.57hr, AD-MT 1.95hr, ALHVR 3.41hr, SKTC-Net 2.94hr. Note that this is not an additional experiment; it is a numerical count of an existing experiment.
Cosine Similarity in RSC (R3): Chosen for scale-invariance across layers/students, angular alignment capturing topological relationships for anatomical clustering, and simplified gradient computation via L2-normalized vectors.
Discussion on SAC (R3): Park et al. addresses debiasing in classification via bias-contrastive pairs and Grad-CAM. SAC serves a fundamentally different purpose in semi-supervised segmentation: it enforces dual-student consensus on spatial saliency maps to localize anatomical regions in unlabeled medical images, without auxiliary samples or bias-negative scoring. The technical forms differ: Park et al. uses dot-product similarity with relative-exploitation scores; SAC uses channel-wise power aggregation with L2 normalization.
Notation, Readability and Missing Related Works (R1, R3): We acknowledge these issues and will fix them: reserve l for layer index and D_L for labeled data; explicitly define L_con as standard semi-supervised consistency loss (Dice between predictions, CPS-style); restructure long sentences. These are presentation improvements. Due to space constraints, the submitted version had only limited space for literature discussion. We will incorporate these recent works in future versions to refine the literature review section.
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.
This paper received mixed reviews, with major concerns focusing on comparisons with similar works and the claim of “mimicking clinical reasoning.” Also, the technical contribution is challenged since it combines several existing designs without a clear justification. In addition, the theoretical relevance between the model design and 2D/3D anatomical structures requires further clarification. Therefore, the authors are invited to submit a rebuttal.
- 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.
After the rebuttal, all reviewers provided positive comments on this work. The main contribution is a “knowledge-level” consistency training beyond the common pixel-level regularization. Experiments demonstrated its effectiveness. The authors should further refine the paper in the camera-ready version, and a final acceptance recommendation is given.
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.
After carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper proposes SKTC-Net, a semi-supervised medical image segmentation framework that extends consistency learning beyond pixel-level predictions to structured representations, including spatial attention, inter-sample relations, and inter-layer feature transitions. Although the novelty is somewhat incremental and the clinical reasoning motivation may be more conceptual than mechanistic, the method is technically sound, clearly motivated, and empirically effective across multiple benchmarks. The rebuttal adequately addressed the main concerns regarding 3D applicability, computational overhead, the distinction between SAC and prior attention methods, the use of cosine similarity in RSC, and presentation issues such as notation ambiguity and missing definitions. Given the consistent performance improvements, compatibility with existing semi-supervised frameworks, and unanimous post-rebuttal support from the reviewers, I recommend acceptance, with the expectation that the authors will fully implement the promised clarifications and readability improvements in the final version.
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.
The proposed structured knowledge topology consistency network (SKTC-Net) is recognized as a well-motivated and empirically effective framework for semi-supervised medical image segmentation. While initial concerns were raised regarding the incremental nature of the modules, 3D generalizability, and presentation ambiguities, consensus among reviewers post-rebuttal highlights that the authors have resolved these concerns through convincing empirical clarifications and committed manuscript revisions.