List of Papers Browse by Subject Areas Author List
Abstract
Tissue-level semantic segmentation is vital to computational pathology, yet pixel-wise annotation of whole-slide images remains prohibitively expensive, motivating weakly supervised approaches. Existing weakly supervised semantic segmentation methods face two coupled challenges: extreme intra-class morphological heterogeneity and pseudo-label noise accumulation, which jointly induce progressive semantic drift during training. In this paper, we propose PhysPro, a physics-semantics synergistic framework which stabilizes weak supervision by anchoring representation learning to immutable physical priors. Specifically, PhysicsDerived Tri-Domain Static Prototypes (PhyTriSP) construct offline frozen anchors across the semantic, phase-morphological, and spectral-energy domains, which provide physically grounded references that mitigate intra-class variability. Building upon these anchors, Physics-Semantics Orchestrated Mixture of Experts (PhyOrchMoE) deterministically decouples spatial morphology from spectral energy via physics-steered triscale fusion and sub-band amplitude calibration, which yields complementary noise-insulated representations. A Physics-Guided Discrepancy Supervisor (PhyDS) then exploits inter-domain discrepancy across the frozen anchors, which disentangles label noise from tissue boundary ambiguity to enable tri-regime adaptive supervision with prototype-guided soft labels and category-aware spectral recalibration. Extensive experiments on BCSS-WSSS and LUAD-HistoSeg demonstrate that PhysPro consistently outperforms state-of-the-art methods.The code is available at https://github.com/zsyUp/PhysPro.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1322_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zsyUp/PhysPro
Link to the Dataset(s)
BCSS-WSSS: https://drive.google.com/drive/folders/1iS2Z0DsbACqGp7m6VDJbAcgzeXNEFr77
LUAD-HistoSeg: https://drive.google.com/drive/folders/1E3Yei3Or3xJXukHIybZAgochxfn6FJpr
BibTex
@InProceedings{ZhaShu_PhysicsGrounded_MICCAI2026,
author = { Zhang, Shuyu AND Fan, Jiansong AND Mao, Wanglong AND Li, Lihua AND Jiang, Yizhang AND Qian, Pengjiang AND Pan, Xiang},
title = { { Physics-Grounded Weakly Supervised Histopathological Tissue Segmentation via Frozen Tri-Domain Prototypes } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes PhysPro, a weakly supervised histopathology tissue segmentation framework that aims to stabilize training under severe intra-class heterogeneity and noisy pseudo-labels. The method constructs frozen prototypes in three domains (semantic, phase-morphological, and spectral-energy), uses them to guide a deterministic morphology/frequency mixture-of-experts module, and further leverages cross-domain discrepancy to distinguish reliable pixels, boundary pixels, and noisy pseudo-labels for adaptive supervision. The reported results on BCSS-WSSS and LUAD-HistoSeg are competitive and outperform the baselines included in the paper.
- 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 addresses a real and important pathology-specific challenge: weak supervision is particularly fragile in histopathology due to strong morphological diversity and pseudo-label noise. 2.The overall framework is coherent, and the three components are reasonably well connected at the conceptual level. 3.The use of phase and amplitude information is intuitively meaningful in this domain, and the empirical results are competitive on the reported benchmarks.
- 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 methodological novelty appears moderate rather than strong. While the exact combination is new, many core ingredients already have clear precedents in recent literature [7, 8, 12, 20, 21], including prototype-based guidance, prototype refinement/alignment for WSSS, and noise-aware pseudo-label handling. As a result, the paper reads more like a sophisticated recombination of existing ideas than a fundamentally new formulation.
2.A key methodological ambiguity is how the category-specific sets used for prototype construction are formed under weak supervision. Since only image-level labels are available and pathology patches may contain multiple tissue types, it is unclear how clean semantic, phase, and energy prototypes are obtained without substantial contamination from off-class regions. This point is central to the method, yet it is insufficiently specified.
3.The discrepancy supervisor is under-defined. The paper argues that cross-domain discrepancy can separate reliable pixels, true boundary ambiguity, and corrupted pseudo-labels, but the actual decision rule for assigning pixels to these regimes is not clearly described. Since pseudo-label noise handling is a major challenge in WSSS, the thresholding and regime-definition details are important for assessing soundness and reproducibility.
4.The paper makes strong mechanism-level claims that are not fully validated experimentally. For example, it repeatedly argues that frozen prototypes prevent semantic drift and that discrepancy disentangles boundary ambiguity from label noise, but these claims are supported mainly by overall performance gains and a limited ablation table, rather than direct evidence such as drift analysis or frozen-vs-learnable prototype comparisons.
5.The experimental section is somewhat thin for a method with several interacting components. Important ablations are missing, such as frozen vs. learnable prototypes, phase-only vs. energy-only variants, and analyses of the necessity of all three domains.
6.The fairness of comparison is not fully clear because the method uses MedCLIP-derived semantic prototypes, while it is unclear whether competing baselines are given access to comparably strong pretrained vision-language priors. Since MedCLIP is itself a strong medical vision-language pretraining model, this should be clarified and ideally isolated in ablation.
7.The “physics-grounded invariance” motivation is intuitively appealing, but it is not directly validated with robustness experiments. If the main argument is that phase and spectral-energy priors improve robustness to staining variation, appearance variation, or scanner/domain shifts, then targeted perturbation or cross-domain experiments would be important. In the current version, this claim remains more intuitive than empirically demonstrated.
- 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?
The paper is well motivated and technically coherent, and the reported results are promising. I appreciate the attempt to address a real pathology-specific challenge in weakly supervised segmentation by combining frozen tri-domain prototypes, physics-aware feature decoupling, and discrepancy-based adaptive supervision. However, my overall score was mainly limited by three factors: the methodological novelty appears moderate rather than substantial, several core implementation details are under-specified, and the experimental validation is not yet deep enough to fully support the stronger mechanism-level claims made in the paper. In particular, the paper does not clearly explain how prototypes are reliably constructed under weak supervision, how the discrepancy-based regimes are defined in practice, or provide enough direct analysis to validate claims about semantic-drift prevention and boundary/noise disentanglement. Overall, I find the paper interesting and potentially valuable, but not yet sufficiently convincing in its current form.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I thank the authors for their careful rebuttal and for the additional clarifications. The response partially addresses my concerns, especially by clarifying how the prototype sets are constructed and by providing more explicit decision rules for the PhyDS regime assignment.
However, my main concerns are only partially resolved. The novelty concern remains, as the rebuttal mostly restates the intended role of frozen tri-domain prototypes rather than providing stronger evidence that this constitutes a substantially new formulation beyond a combination of prototype guidance, frequency-domain priors, and noise-aware supervision. More importantly, the rebuttal does not provide key mechanism-level analyses that would be needed to fully substantiate the stronger claims, such as frozen vs. learnable prototypes, semantic/phase/energy single-domain variants, direct validation of boundary/noise separation, or robustness under stain/scanner/domain shifts. The authors rely mainly on the existing component-level ablation, which is helpful but not sufficient to validate the specific claimed mechanisms.
Overall, I appreciate the authors’ effort and acknowledge that the rebuttal improves the clarity of the method. Nevertheless, the main issues regarding novelty, mechanism-level validation, and robustness evidence remain insufficiently addressed. Therefore, I maintain my recommendation to reject.
Review #2
- Please describe the contribution of the paper
This paper presents PhysPro, a physics-semantics synergistic framework designed for weakly supervised semantic segmentation in histopathology. The authors address the core challenges of extreme intra-class morphological heterogeneity and the accumulation of pseudo-label noise by anchoring representation learning to immutable physical priors. Experiments show improved segmentation performance.
- 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.
Proposes a novel perspective by using frequency-domain representations as physically-grounded descriptors to stabilize weak supervision.
Demonstrates superior performance over existing methods on two large-scale benchmarks, BCSS-WSSS and LUAD-HistoSeg, with clear gains in mDice and mIoU.
- 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.
Although the authors claim the frequency-domain features are robust to staining fluctuations, the experiments are limited to two specific datasets; broader validation across more diverse scanning protocols and multi-center data would further strengthen the claims of physical invariance
- 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 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 method design for using frequency-domain representations as physically grounded descriptors to stabilize weak supervision.is interesting, and also shows improved performance across benchmarks.
- 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 author has appropriately addressed the main questions I raised in the first round, so I have moved from “weakly accept” to “accept”.
Review #3
- Please describe the contribution of the paper
This paper proposed PhysPro, a physicssemantics synergistic framework which stabilizes weak supervision by anchoring representation learning to immutable physical priors. Specifically, Physics Derived Tri-Domain Static Prototypes (PhyTriSP) construct offline frozen anchors across the semantic, phase-morphological, and spectral-energy domains, which provide physically grounded references that mitigate intra-class variability. Building upon these anchors, Physics-Semantics Orchestrated Mixture of Experts (PhyOrchMoE) deterministically decouples spatial morphology from spectral energy via physics-steered triscale fusion and sub-band amplitude calibration, which yields complementary noise-insulated representations. A Physics-Guided Discrepancy Supervisor (PhyDS) then exploits inter-domain discrepancy across the frozen anchors, which disentangles label noise from tissue boundary ambiguity to enable tri-regime adaptive supervision with prototype-guided soft labels and category-aware spectral recalibration. Experiments on BCSS-WSSS and LUAD-HistoSeg demonstrate that PhysPro consistently outperforms state-of-the-art methods. Code will be released upon publication.
- 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.
- It is an interesting idea to propose a physicssemantics synergistic framework which stabilizes weak supervision by anchoring representation learning to immutable physical priors.
- The results seem good as shown in Table 1 and Fig. 2.* Ablation analysis verified the efficiency of each component.
- 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.
- It would be good to proof the efficiency of PhysPro mathematically, which would help readers to know its merits and how to use it in the future.
- It is suggested to show the evidence to show why the proposed mothod works for the given cases.
- I am also happy to know whether the proposed methed would work for the natural iamges besides medical images. It is better to apply this method for other cases.
- 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?
It is good to release the code for the readers in this domian. For other weekness, please refer to Q7.
- 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
Response to Reviewer 1: We sincerely thank the reviewer for the thorough and constructive comments. Q1.Novelty. We understand the reviewer’s concern regarding novelty. PhysPro is not a simple recombination of prototypes, frequency-domain priors, and noise-aware supervision; it redefines the role of prototypes in WSSS. Instead of using prototypes only as semantic alignment targets, PhysPro builds frozen tri-domain references before training. The semantic, phase-morphological, and spectral-energy anchors encode class identity, structural consistency, and texture-frequency distribution, respectively. Their discrepancy is used as a diagnostic signal for supervision reliability, helping separate reliable interiors, boundary ambiguity, and corrupted pseudo-labels. This diagnostic use of frozen tri-domain prototypes differs from conventional prototype-based WSSS. Q2.Prototype Construction. S_c is built from training patches whose image-level label contains only a single tissue category; multi-category patches are excluded. No CAMs, pseudo-labels, or segmentation predictions are used, so construction is decoupled from training dynamics. We do not assume pixel-level purity; this subset provides a lower-contamination calibration pool, while Eq. (1) further suppresses residual off-class structures/outliers by density-weighted aggregation. Patch counts range from 679 to 4,738 per category on BCSS-WSSS. Q3.Decision Rules of PhyDS. For each pixel with pseudo-label c, PhyDS computes s_sem, s_phase, and s_energy; Delta = Var(s_sem, s_phase, s_energy). Confidence q combines mean tri-domain similarity with 3x3 local consistency using weights 0.7/0.3.Thresholds are batch-adaptive: 30th/70th percentiles for low/high Delta, and medians for high q and s_sem. Assignment: (1) low Delta, high q: reliable; (2) high Delta, high s_sem: boundary ambiguity; (3) high Delta, low s_sem: corrupted pseudo-label; (4) otherwise: attenuated supervision. Hyperparameters are beta=0.5, tau=0.1, T=0.5.We will add these details to Sec. 2.4.Q4/Q5/Q7.We thank the reviewer for the helpful suggestions. Table 2 provides component-level evidence: PhyTriSP yields the largest gain, and the full model further improves with PhyOrchMoE/PhyDS, supporting the contribution and complementarity of the modules. Finer single-domain variants can further disentangle the roles of semantic, phase, and energy priors; our tri-domain design is motivated by their complementarity rather than interchangeable comparison. Consistent gains on two public WSSS pathology benchmarks from different cancer types and data sources further support the effectiveness of this design. Q6.Fairness of MedCLIP. PBIP, the closest prototype baseline, also uses MedCLIP for prototype construction, so the comparison is not against a prototype method without comparable medical VLP priors. In PhysPro, MedCLIP only constructs the frozen offline P_sem; it is not the backbone and produces no dense predictions.
Response to Reviewer 2: We sincerely thank the reviewer for recognizing the frequency-domain perspective and for the constructive suggestion. Broader multi-center and multi-protocol validation would further strengthen the physical-invariance analysis. In the current study, consistent improvements on two histopathology benchmarks provide empirical support for PhysPro across different dataset settings and support the effectiveness of its frequency-domain design.
Response to Reviewer 3: We sincerely thank the reviewer for the encouraging assessment and thoughtful suggestions. Phase congruency is stable to monotonic intensity transformations, while log-polar angular integration reduces rotational dependence. These analytic properties support prototype stability. Together with semantic/phase/energy heterogeneity, they help Delta distinguish boundary ambiguity from label noise, as supported by Table 2 and Fig. 2.These general signal-processing principles also suggest potential applicability to natural-image WSSS.
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 paper proposes a physics-grounded framework for weakly supervised histopathological segmentation. The reviewers’ scores are 3, 4, and 4, indicating mixed opinions. The reviewers agree that the problem is relevant and the proposed framework is interesting, with promising empirical results. However, concerns remain regarding the level of methodological novelty, insufficient clarification of key components, and limited experimental validation of the underlying claims. The AC therefore recommends inviting the paper for rebuttal.
The authors are encouraged to clarify the construction and reliability of the proposed prototypes under weak supervision, provide stronger validation of the claimed mechanisms, and further justify the experimental design and comparisons.
- 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 PhysPro, a weakly supervised histopathology tissue segmentation framework that uses frozen prototypes from three domains: semantic, phase-morphological, and spectral-energy. These prototypes are used to guide a morphology/frequency mixture-of-experts module and a cross-domain discrepancy-based supervision strategy, with the goal of improving robustness to intra-class heterogeneity and noisy pseudo-labels in weakly supervised tissue segmentation.
The initial reviews were mixed but overall slightly positive, with scores of 3 / 4 / 4.The reviewers appreciated the clinically relevant problem, the coherent design of the three-domain prototype framework, and the competitive results on BCSS-WSSS and LUAD-HistoSeg. However, several concerns were raised. Reviewer 1 questioned the degree of methodological novelty, arguing that the method appeared to recombine existing ideas such as prototype-guided learning, prototype refinement/alignment, and noise-aware pseudo-label handling. Additional concerns included insufficient clarity on how category-specific prototype sets are constructed under weak supervision, how the discrepancy supervisor assigns pixels to reliable/boundary/noisy regimes, limited mechanism-level evidence for the claimed role of frozen prototypes, and incomplete ablations and fairness analysis.
The rebuttal addressed several of these points. The authors clarified that PhysPro is intended to redefine the role of prototypes in weakly supervised segmentation by constructing frozen tri-domain references before training, rather than using prototypes only as semantic alignment targets. They also explained how prototypes are constructed from single-label training patches, how the discrepancy decision rules are defined, and how the method differs from MedCLIP-based prototype baselines. They further clarified the role of the three components and the frequency-domain design, and argued that the frozen semantic, phase-morphological, and spectral-energy anchors provide complementary information.
Some concerns remain. In particular, Reviewer 1 was only partially persuaded by the rebuttal and maintained that the novelty and mechanism-level evidence remain limited. The rebuttal improves the framing and clarifies several design choices, but it does not fully remove the concern that the method is a sophisticated combination of existing weakly supervised segmentation components. The final version should more clearly explain the prototype construction procedure, discrepancy-thresholding rules, and the specific evidence supporting the claimed role of frozen tri-domain prototypes.
After rebuttal, the reviewer opinions are effectively R / A / A, treating the no-update weak accept reviewer as retaining the initial positive assessment. Reviewer 1 maintained reject because the novelty and mechanistic validation concerns were only partially addressed. Reviewer 2 recommended accept after the rebuttal, noting that the main questions were addressed. Reviewer 3 initially recommended weak accept and did not provide a post-rebuttal update, so I interpret that review as remaining positive.
I recommend Accept. The paper remains somewhat borderline because the methodological novelty is not overwhelming and some mechanism-level evidence should be strengthened. However, the overall reviewer balance is positive, the rebuttal clarified several important design choices, and the paper addresses a relevant weakly supervised histopathology segmentation problem with a coherent framework and competitive empirical results. The remaining issues are important limitations, but I do not consider them fatal. The camera-ready version should clarify the novelty framing, prototype construction, discrepancy assignment rules, and remaining limitations.
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 reviewers have quite different opinions for this paper. One reviewer is mainly concerned about novelty and validation of the method. While empirically the results are good on two commonly used dataset, the physics-aware aspects are more heuristic rather than formally formulated or validated. The overall method is also quite complex
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 rebuttal, this paper receives two accept recommendations and one rejection. The AC acknowledges that PhysPro addresses an important weakly supervised histopathology segmentation problem and proposes an interesting physics-semantics framework using frozen tri-domain prototypes across semantic, phase-morphological, and spectral-energy domains. The rebuttal clarifies several key methodological details, including prototype construction from lower-contamination single-category patches, batch-adaptive PhyDS decision rules, the role of MedCLIP, and the rationale for using cross-domain discrepancy to guide adaptive supervision. Although concerns remain regarding moderate novelty, limited mechanism-level validation, and the need for broader robustness studies under stain/scanner/domain shifts, the method is coherent, empirically competitive on two public benchmarks, and positively assessed by the majority of reviewers. Therefore, the AC recommends acceptance.