List of Papers Browse by Subject Areas Author List
Abstract
Accurate tumor microenvironment (TME) analysis is critical for prognosis but limited by the high cost of pixel-level annotations. Current weakly supervised methods suffer from localization bias, and although semi-supervised segmentation yields superior boundaries, standard frameworks depend on pixel-level annotations. To resolve this, we propose Homogeneity driven Segmentation, namely HomoSeg, a weak-label-guided semi-supervised framework. It exploits histological spatial homogeneity to extract single-label samples. Superpixel segmentation is utilized to synthesize training samples with dense and reliable pixel-level labels. Subsequently, we design a multi-task Siamese network integrating classification, segmentation, and CAM branches. Through weak-to-strong consistency regularization, supervision signals are propagated from synthetic data to unlabeled data, establishing robust semantic decision boundaries. Finally, to eliminate noise from synthetic data and achieve model lightweighting, we employ an offline distillation strategy to distill the multi-task Siamese network. Experimental results on the LUAD-HistoSeg and BCSS-WSSS datasets demonstrate that HomoSeg significantly outperforms existing methods, providing a robust solution for high-precision Tumor Microenvironment analysis. The code is available at: https://github.com/zhy1312/HomoSeg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2263_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhy1312/HomoSeg
Link to the Dataset(s)
LUAD-HistoSeg dataset: https://drive.google.com/drive/folders/1E3Yei3Or3xJXukHIybZAgochxfn6FJpr
BCSS-WSSS dataset: https://drive.google.com/drive/folders/1iS2Z0DsbACqGp7m6VDJbAcgzeXNEFr77
BibTex
@InProceedings{ZhaHao_HomoSeg_MICCAI2026,
author = { Zhao, Haoyun AND Ni, Jun AND Zhan, Yibing AND Qi, Chongchong AND Zhu, Yuhe AND Tao, Dapeng},
title = { { HomoSeg: A Weak-Label-Guided Semi-supervised Framework for Tumor Microenvironment 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
The paper proposes HomoSeg to address the prohibitive cost of pixel-level annotation in tumor microenvironment segmentation. The core logic hinges on the spatial homogeneity prior of histopathological images, founded on the observation that a local patch cropped from a Whole Slide Image typically contains a single histological category. By leveraging this prior, the framework transforms image-level weak labels into reliable pixel-level supervisory signals, thereby reformulating the weakly supervised task into a semi-supervised learning framework.
- 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 proposes a data reconstruction strategy that formalizes the pathological prior that WSI-cropped patches tend to contain a single tissue category, enabling supervised training subset construction without any pixel-level annotation. The paper introduces a superpixel-based synthesis strategy that improves boundary quality over standard CutMix, where irregular boundaries better align with real tissue morphology and directly address the most challenging boundary regions in TME segmentation. The paper incorporates an offline distillation strategy that substantially reduces training cost without compromising segmentation accuracy, offering practical reference value for clinical deployment scenarios.
- 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 paper lacks quantitative evidence to support the homogeneity hypothesis. The authors do not report the actual proportion of single-label patches within the training set, making it impossible to assess the validity of this prior across the utilized datasets or the sensitivity of the proposed method to this assumption. The paper exhibits high architectural similarity to HisynSeg, as both leverage image synthesis and consistency regularization for weakly supervised pathology segmentation. However, the manuscript fails to provide a targeted comparative analysis to clarify the fundamental methodological distinctions between the two frameworks. The paper adopts a pseudo-label generation strategy for CAM calibration that simply computes an unweighted average of CAM activations and segmentation predictions. This approach fails to model the confidence disparity between the two branches, potentially introducing noise during early training stages when both components are unstable. Notably, the paper provides neither analytical discussion nor ablation studies to justify this design. The paper lacks a discussion on the application boundaries of the proposed method. Its performance remains unverified in scenarios involving rare tissue categories or extreme class imbalances, leaving the limits of its generalization capability unclear.
- 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 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed data reconstruction approach, driven by the spatial homogeneity prior, demonstrates domain-specific relevance, and the experimental results show performance improvements over baseline methods across two public datasets. However, the core hypothesis lacks quantitative validation, the fundamental distinctions from recent works such as HisynSeg are not clearly defined. Moreover, the design of the pseudo-label generation strategy is oversimplified and lacks targeted ablation, and the application boundaries and failure scenarios of the method are entirely undiscussed.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the rebuttal improves clarity, the core concerns regarding the validity of the homogeneity assumption, the methodological distinction from prior work, and the lack of targeted validation for the pseudo-label strategy remain insufficiently resolved; therefore, I maintain my recommendation for rejection.
Review #2
- Please describe the contribution of the paper
This paper proposes HomoSeg, a weak-label-guided semi-supervised framework for tumor microenvironment segmentation. It first generates synthetic data with dense annotations from weak labels, then uses a semi-supervised pipeline to train the segmentation network, and finally applies distillation to obtain a clean segmentation model. The authors validate the method on the LUAD-HistoSeg and BCSS-WSSS datasets, where it outperforms existing methods.
- 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 authors propose a novel data reformulation strategy based on the observation that all pixels in a patch can share the same label when the patch contains only a single tissue type. This strategy is useful for converting weak labels into dense labels and further supports synthetic data generation. 2.The HomoSeg framework achieves satisfying experimental performance.
- 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 proposed method has a fairly heavy design with many components, while each individual component brings only marginal gains, as shown in Table 2.In particular, synthetic data generation appears to be a substantial part of the claimed contribution, but removing it leads to only limited performance differences, especially on LUAD-HistoSeg.
2.Regarding the semi-supervised part: the labeled/unlabeled split with a Siamese network, strong/weak augmentations, multi-task co-training, using CAM for pseudo-label generation, and the loss designs all seem to be fairly common practices in semi-supervised learning from recent years. Rather than simply stacking these components together, could the authors clarify what the main distinguishing idea of this part is?
3.Regarding the distillation part: the paper does not specify how the student model is initialized, and I assume it is trained from scratch. The supervision signal is derived from the averaged CAM and teacher predictions, so its performance should in principle be bounded by that signal unless there is some additional design that enables further improvement. Why not directly use the combined teacher and CAM predictions instead of retraining a student model? I am also confused by the comparison in Table 3.My understanding is that the teacher model corresponds to the full pipeline in Stage II, while the student model is the distilled encoder-decoder segmentation model. However, the student model cannot exist unless the teacher model has already been trained. In that case, distillation does not reduce the end-to-end training time or memory cost, and the reported inference time also shows no difference.
4.Some presentation details also need clarification. In Figure 1, please specify what “SSN” stands for, either in the figure or in the main text. In Equations (2) and (3), does the symbol denotes multiplication? In Table 1, please also define the abbreviations such as TE and NEC.
- 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?
Despite several weaknesses that require further clarification and reconsideration, the overall framework proposed by the authors is useful, and some parts of the method offer new insights. Overall, I would give this paper a weak accept.
- 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 pipeline still looks over complicated. But the reformulation of solving weakly by semi is acceptable.
Review #3
- Please describe the contribution of the paper
HomoSeg proposes a weak-label-guided semi-supervised framework for tumor microenvironment (TME) segmentation in histopathology images. The core insight is that WSI-cropped patches often contain only a single tissue category, making image-level labels locally equivalent to pixel-level annotations. This “spatial homogeneity” prior is exploited in three stages: (1) Homogeneity Driven Data Reformulation, which mines single-label samples and synthesizes multi-label training data via superpixel-based CutMix to create a labeled set with dense pseudo-masks; (2) a multi-task Siamese network (FixMatch-based) that jointly trains classification, segmentation, and CAM branches with weak-to-strong consistency regularization to propagate supervision from synthetic labeled data to unlabeled real data; and (3) an offline distillation strategy that uses the converged teacher network to generate a clean labeled dataset for training a lightweight student model, reducing computational overhead and eliminating synthetic noise. Experiments on LUAD-HistoSeg and BCSS-WSSS demonstrate state-of-the-art performance under weak supervision.
- 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) Elegant exploitation of a domain-specific prior. The observation that WSI patches tend to be spatially homogeneous and thus single-label patches yield reliable global pixel masks, is both biologically motivated and practically effective. This insight elegantly bridges the gap between weak and semi-supervised learning without any additional annotation cost. 2) Well-integrated three-stage pipeline. The three components (data reformulation, semi-supervised training, distillation) are logically coherent and each addresses a distinct limitation of the prior stages. The CAM-calibrated pseudo-label fusion (Eq. 7) is a sensible design to handle early-training instability in the segmentation branch. 3) Computational efficiency gains from distillation. Table 3 shows a dramatic reduction in training cost (852 > 217 s/epoch, 17.89 > 5.74 GB GPU memory) with simultaneous accuracy improvement, which is practically very relevant for clinical deployment.
- 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) Single-label proportion not reported. The central assumption of the method, that a meaningful proportion of training patches are single-label, is never quantified for either dataset. If only a small fraction of samples satisfy ||y||₁ = 1, the homogeneity-driven reformulation may be of limited benefit, and the method may not generalize to datasets with different WSI cropping strategies or tumor types. 2) No comparison with pixel-supervised upper bounds. The paper does not include any fully supervised baseline to provide an upper bound on performance. Without this, it is unclear how much of the performance gap remains relative to the “ideal” scenario, which would better contextualize the practical clinical utility. 3) Limited discussion of failure cases. The qualitative results in Fig. 2 show only relatively favorable examples. There is no analysis of where HomoSeg fails (e.g., rare tissue categories, highly ambiguous boundaries, samples that violate the homogeneity assumption), which would help readers understand the method’s limitations. 4) Synthetic data realism is not validated. The superpixel-based CutMix synthesis creates composite images with artificial tissue boundary arrangements that do not occur in real histology. There is no qualitative or quantitative analysis of how realistic or harmful these synthetic boundaries are. The claim that they are “morphologically realistic” is not substantiated beyond the method description.
- 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 primary factors driving the score are: (1) the novelty and elegance of the spatial homogeneity prior as a bridge between weak and semi-supervised learning; (2) consistent and meaningful improvements on two public benchmarks with a thorough ablation; and (3) the efficiency gains from the distillation stage, which add practical value.
- Reviewer confidence
Somewhat confident (2)
- [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 clarified the experimental details requested and addressed other raised concerns.
Author Feedback
We sincerely thank the Meta-Reviewers and reviewers for your constructive feedback. We address core concerns below, with all evidence strictly from the submitted manuscript. We systematically clarify the key issues: 1.Homogeneity Hypothesis Validation & Application (R1Q1/Q4, R3Q1, Meta) Exact dataset statistics: 27.55% of LUAD-HistoSeg and 40.04% of BCSS-WSSS training patches are single-label, confirming spatial homogeneity is prevalent in WSI cropping. Our method does not require majority single-label samples. The mined subset only serves as initialization anchors, with weak-to-strong consistency propagating supervision to all multi-label samples. The submitted results demonstrate robustness to extreme class imbalance: HomoSeg achieves 77.32% mIoU on the 4.18% rare LYM category and outperforms all SOTA methods on the minority NEC category. 2.Methodological novelty of HomoSeg (R1Q2, R2Q2, Meta) The core contribution of HomoSeg is not simply stacking semi-supervised modules, but reformulating strict weakly supervised tissue segmentation into a semi-supervised paradigm without manual pixel-level annotations. Existing WSSS methods suffer from localization bias, while conventional SSSS frameworks require expensive pixel-level labels. HomoSeg bridges these paradigms by automatically constructing a reliable labeled subset through spatial homogeneity mining. Furthermore, compared to HisynSeg, HomoSeg differs fundamentally: (1)Framework: HisynSeg combines disjoint fully/weakly supervised learning; HomoSeg reformulates the task as a unified semi-supervised framework. (2)Strategy: HisynSeg uses mathematical generation (Bézier curves) introducing artificial structures; HomoSeg uses superpixel-based CutMix preserving real biological boundaries. (3)Consistency: HisynSeg only enforces CAM-segmentation consistency; HomoSeg implements weak-to-strong consistency leveraging unlabeled data directly. Results validate these advantages: HomoSeg outperforms HisynSeg by 3.06% mIoU and 1.69%. 3.Pseudo-label Strategy & Component Synergy(R1Q3, R2Q1Q3, R3Q2, Meta) Clearer experimental analysis and interpretation. The moderate performance gain from synthesis on LUAD-HistoSeg mainly arises from its lower proportion of single-label samples, limiting synthesis diversity compared with BCSS-WSSS. Regarding the unweighted CAM-segmentation averaging, we intentionally adopt a simple calibration mechanism to avoid introducing noisy confidence estimation under weak supervision. The averaging acts as a lightweight mutual regularization between localization priors and dense predictions. Furthermore, the student surpasses the teacher because the teacher is optimized under multi-task objectives with dynamic pseudo-label noise, whereas the student learns from static, refined ensemble targets under a single segmentation objective. This offline distillation mainly aims to suppress optimization noise, simplify the optimization pipeline, and reduce computational overhead compared with iterative refinement strategies. Finally, a recent TMI 2025 study showed that advanced weakly supervised pathology segmentation can approach fully supervised performance with only marginal gaps on BCSS-WSSS, justifying current upper bounds. 4.Regarding synthetic realism (R3 Q4) Unlike rectangular CutMix, our superpixel strategy inherently preserves tissue continuity. These synthesized boundaries serve as auxiliary perturbations for consistency regularization rather than direct pixel-level targets, making our morphology-driven approach significantly more histologically plausible than mathematical geometric synthesis. 5.Clarity & Reproducibility (All Reviewers) We sincerely apologize for unclear presentations, and will revise the final manuscript to complete missing abbreviations, clarify symbol definitions, correct grammatical errors, and publicly release the full code and pre-trained models.. Thank you again for your valuable time and feedback, which has greatly improved our manuscript.
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.
Authors should meticulously address reviewers’ concerns. Especially, the concerns about the homogeneity hypothesis, method novelty, more clear experimental result analysis, and critical settings (i.e., proportion of single label), etc.
- 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 reviewers recognize the value and effectiveness of the proposed framework, and the rebuttal addressed several clarification concerns, leading to two positive recommendations.
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.
The paper proposes HomoSeg, a weak-label-guided semi-supervised framework for tumor microenvironment segmentation. The key idea is to exploit the spatial homogeneity prior in histopathology, where a WSI-cropped local patch often contains a single tissue category, to transform image-level weak labels into pixel-level supervision. The framework uses a data reconstruction strategy, superpixel-based synthesis, and consistency/distillation components to reduce the cost of pixel-level annotations for TME segmentation.
The initial reviews were mixed but overall positive, with scores of 2 / 4 / 4.Reviewers 2 and 3 appreciated the practical motivation, the weak-to-semi-supervised reformulation, the attempt to reduce dense annotation burden, and the reported empirical gains over existing methods. Reviewer 1 was substantially more critical and raised concerns about the validity of the spatial homogeneity hypothesis, the lack of quantitative evidence supporting that assumption, the unclear distinction between weakly supervised and semi-supervised formulations, limited comparative analysis against relevant WSSS and semi-supervised methods, insufficient ablation of the core modules, and unclear application boundaries.
The rebuttal addressed several of these concerns. The authors provided dataset statistics supporting the prevalence of single-label patches, clarified that the mined single-label subset is used only as initialization anchors rather than requiring most samples to be single-label, and argued that weak-to-strong consistency propagates supervision to multi-label samples. They also clarified the methodological distinction between HomoSeg, HistoSeg, and HiSynSeg, explaining that HomoSeg reformulates strict WSSS into a semi-supervised framework without manual pixel-level annotation. The rebuttal further explained the role of superpixel-based synthesis, pseudo-label calibration, offline distillation, and consistency regularization, and committed to improving clarity and reproducibility.
Some limitations remain. The spatial homogeneity assumption is better supported after rebuttal, but it remains a dataset- and patch-size-dependent prior that should be discussed carefully. The paper should also more clearly define its application boundaries, especially for highly heterogeneous TME regions. Some of the comparison and ablation concerns remain only partially addressed, and the final version should strengthen the explanation of why each component is needed and where the method may fail.
After rebuttal, the reviewer opinions are R / A / A. Reviewer 1 maintained reject, stating that the rebuttal improved clarity but did not fully resolve concerns about the homogeneity assumption, methodological distinction, and lack of targeted validation. Reviewers 2 and 3 recommended accept, with Reviewer 3 noting that the authors clarified the requested experimental details and addressed other concerns. Reviewer 2 also found the weak-to-semi reformulation acceptable despite noting that the pipeline remains somewhat complex.
I recommend Accept. Although one reviewer remained unconvinced, the majority of post-rebuttal opinions are positive, and the rebuttal substantially improved the framing of the key assumption and methodological contribution. The paper addresses an important annotation-cost problem in tumor microenvironment segmentation and presents a practically useful framework. The remaining concerns should be explicitly discussed in the camera-ready version, especially the limits of the spatial homogeneity prior, application boundaries, and the need for clearer ablation and comparison details.
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.
After carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper proposes HomoSeg, a weak-label-guided semi-supervised framework for tumor microenvironment segmentation in histopathology images. Its main contribution is the use of a spatial homogeneity prior to reformulate image-level weak labels into dense supervisory signals, followed by semi-supervised learning and offline distillation. The reviewers generally recognized this as a useful and domain-relevant idea, with consistent empirical improvements on public pathology benchmarks and practical efficiency gains from distillation. Although concerns remain regarding the quantitative validation of the homogeneity assumption, the distinction from closely related methods such as HisynSeg, and the complexity of the overall pipeline, the majority of reviewers found the rebuttal acceptable and supported acceptance. Therefore, I recommend acceptance, with the expectation that the authors will more clearly quantify the homogeneity assumption, clarify methodological differences from prior work, and discuss failure cases and application boundaries in the final version.