Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

A unified multiplex virtual staining model enables scalable and non-destructive multiplex analysis from H&E slides while promoting parameter efficiency, shared pathological knowledge, and consistent cross-biomarker representations. However, in clinical practice, data for new biomarkers are typically acquired sequentially over time. Fine-tuning on such temporally arriving data leads to severe performance degradation on previously learned biomarkers, as sequential optimization disrupts the structured relationships among biomarker representations in the latent space. To address this issue, we propose ContiStain, an IHC multi-domain relational distillation framework for continual virtual staining. We first (i) construct a domain-aware structured feature space using a mixture-of-experts (MoE) feature extractor to reduce representation interference across biomarker domains. Based on this stabilized feature space, we then (ii) propose a relation-preserving distillation strategy that explicitly enforces the consistency of cross-domain token-level cosine similarity matrices between learned biomarker domains during continual adaptation. By maintaining cross-domain structural coherence, ContiStain mitigates forgetting while retaining adaptability to new domains. Experiments on the MIST dataset under a four-domain sequential virtual IHC staining setting show improved stability, reducing FID and ConchFID by 11.1 and 60.9 compared to sequential fine-tuning, enabling scalable and robust multi-domain virtual staining. Code is released at https://github.com/ccitachi/ContiStain.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0693_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ccitachi/ContiStain

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CheFuq_ContiStain_MICCAI2026,
        author = { Chen, Fuqiang AND Wang, Yifeng AND Wang, Hongpeng AND Zhang, Yongbing},
        title = { { ContiStain: Cross-Domain Relation-Preserving Distillation for Continual Multi-domain Virtual IHC Staining } },
        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

    The authors investigate virtual IHC staining across domains in a continual learning setting, which is an interesting and relevant topic for pathological research. The main idea is easy to follow, and the results appear promising.

  • 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 authors study the continual learning problem in multi-domain virtual IHC staining, which remains relatively underexplored, and propose a relational distillation-based approach.
    • The experimental results show that the proposed method achieves clear improvements over conventional approaches.
  • 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 reviewer’s main concern is that, even when the previous generative model is used as a frozen teacher, the correlation matrix estimated from it may become highly unreliable or even fail entirely, if the incoming domain differs substantially from previous ones. Consequently, the relational distillation objective may be inherently fragile and may even introduce significant bias.

    • During distillation, even if the experts are frozen, updating the router can still lead to inevitable forgetting. More importantly, such forgetting in the router may disrupt the learned cross-domain relationships.

    • The presentation of the Domain-Aware MoE module lacks clarity. In particular, it remains unclear how patch sampling together with a shared projection can effectively capture structural relationships within each domain.

    • The reviewer believes that merely maintaining previously learned cross-domain relationships is not sufficient to fully address forgetting. For each incoming domain, it is also necessary to jointly model its relationships with prior domains. Therefore, the relation distillation should be extended to the consideration of the incoming domain.

    – The experimental comparisons rely on outdated compared methods, with even the most recent one dating back to 2017, approximately nine years ago.

    – Why are the results of the Initial method on Ki67 not reported? Do the authors assume that the performance of the Initial method on Ki67 is identical to that of the proposed method? Reporting the results on Ki67 (the current domain) is important for evaluating the plasticity of the proposed method.

  • 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.

    (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?

    Although this manuscript is the first to investigate continual learning in virtual staining, the proposed method suffers from several critical concerns in methodology and is not supported by sufficiently comprehensive experimental validation (e.g., outdated compared methods). Therefore, I do not believe the paper currently meets the acceptance bar for MICCAI.

  • 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.

    The reviewer would like to thank the authors’ response. However, my main concerns remain insufficiently addressed.

    • Regrading the relational distillation, even if the distillation is only performed among old domains, the reference relations are still conditioned on new-domains, which can make the correlation matrix unstable and even biased.

    • For router forgetting. Actually the router is updated, and it can directly affect expert selection, changes in routing may still alter old-domain behaviors. Moreover, no empirical analysis has been provided to support the author’s claim for the router stability.

    • For new-old domain relations. The reviewer still believes that preserving only old-old relations may be insufficient, especially limiting the plasticity on the new domains, despite with potential forgetting reduction.



Review #2

  • Please describe the contribution of the paper

    The authors’ propose a continual learning-based multi-domain IHC virtual staining approach where a unified model learns to generate samples of new domains while not forgetting the previous ones. This is the first work to introduce the continual learning problem in virtual staining. Previous works in other domains involve replay-based approaches or distillation-based approaches however they fail to preserve cross-biomarker structural relations. Different biomarkers have different staining patterns but they have the same underlying tissue structure. The proposed approach, ContiStain, has two components: 1) it has a domain-aware feature extractor using MoE, and 2) a relational distillation strategy preserving cross-domain structural consistency within the MoE latent space. The approach uses ASP (a previously introduced virtual staining GAN) at its core and builds on it. Experiments show that ContiStain mitigates forgetting and performs well in the continual learning setup as compared to some prior works.

  • 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.This is the first work to introduce multi-domain continual learning-based virtual staining. 2.The proposed domain-aware feature modeling strategy helps with disentangling domain characteristics from morphological information. 3.The introduced relational distillation strategy further helps with preserving the structural consistency among different domains.

  • 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.It’s not clear whether all continual learning-based baselines use ASP as the starting framework. 2.The authors’ randomly sample P spatial tokens to get structural descriptors, how are these P tokens selected? And why are these P tokens spatial descriptors? It’s also not mentioned what P is set to. 3.According to the authors, the projection function applied to spatial tokens emphasizes morphology-related structures in the unified latent space. How is this disentangling done? I don’t think they are using any loss to make sure that this happens. 4.The authors nowhere mention what the different experts are, and if they are trained end-to-end or are some pre-trained networks. Also, the gating network used to assign weights has also not been discussed. 5.The authors mainly use PSNR/SSIM/DISTS to measure similarity with ground-truth but all of these metrics have been shown to not capture the semantic similarities. ConchFID helps with assessing how close the generated images are to the IHC distribution, however, it’s not a paired metric and only captures distributional similarities. 6.In many cases, the performance of prior works is higher on certain metrics however there’s no analysis on why this might be happening.

  • 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.

    (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?

    It’s not fully clear how some parts in the pipeline work. Furthermore, the evaluations are also not strong enough both in terms of the metrics being used and also the proposed method doesn’t outperform prior works on the reported metrics in several cases. So it’s not clear how good the method actually performs.

  • 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 most of my concerns, and it looks like an interesting approach for continual learning in virtual staining, although I would have liked to see more semantics based metrics being evaluated. If the paper is accepted, the authors should atleast add more details to the method part to make it clear.



Review #3

  • Please describe the contribution of the paper

    This paper proposes applying continual learning to multi-domain virtual IHC staining. The authors first construct a MoE-based encoder to reduce the interference caused by joint learning across different domains, thereby preserving the structural relationships among domain-specific features. This structure is then explicitly maintained when learning from new domains through a relation-preserving distillation loss. The overall approach is reasonable and well-designed, and the experimental results are generally strong.

  • 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.This paper is well-motivated, and the proposed method reasonably addresses continual learning from a feature-level perspective. This is a timely example of applying trending continual learning techniques to the biomedical domain. 2.The writing is clear. The paper is well-structured and easy to follow. 3.Both quantitative results and qualitative visualizations effectively demonstrate the effectiveness of the proposed method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The authors could clarify why the shared feature space in Section 2.2 is said to emphasize structures. Is this property driven by a specific loss that enforces appearance invariance, or does the model simply learn this preference implicitly?

    2.The experimental setting in this paper only considers continual learning over four domains (HER2→ER→PR→Ki67), which is limited in scale. My concern is: 1) Since Eqs. 9, 10 are directly related to the number of domains, increasing the number of domains may lead to a quadratic growth in the number of loss computations. However, the paper does not report the computational cost. In scenarios with many domains (e.g., 10), would the computational overhead become prohibitive? 2) Similarly, in Fig. 2, the pipeline following the Gen Model needs to process more \hat{y} as the number of domains increases. Does this significantly increase the computational cost? 3) As the number of domains continues to increase, can the proposed method still maintain a stable structure in the feature space?

    3.Minor: The symbol \mathcal{L} used in Eq. 8 and on page 4 shares the same notation as the loss, but represents a different meaning. It should be changed. In Fig. 2 Domain-aware MoE Extractor, the symbol G is not explained in either the legend or the main text.

  • 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?

    I consider this to be a solid submission to MICCAI. The method is technically sound, and its application in the biomedical domain is both reasonable and effective. My concerns mainly revolve around the lack of certain clarifications and the absence of discussion on computational scalability when scaling to a larger number of domains. If the authors can provide a clear explanation for these issues, I would be willing to further increase my score.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I thank the authors for their rebuttal. My concerns have been well addressed.

    I am also aware that my rating appears higher than those of the other reviewers. However, after carefully reading the paper and the other reviewers’ comments, I still believe this work is interesting due to its new task formulation and potential practical value. This appears to be a worthwhile direction to explore. In addition, the authors’ responses to the other reviewers were generally convincing to me.

    For these reasons, I will maintain my positive rating.



Author Feedback

R1 Q1: The frozen old model only generates old-domain stains from the same HE input to compute intra-old-domain relations, providing a stable reference independent of the incoming domain. Relational distillation does not directly distill between new and old domains; it only aligns teacher/student old-domain correlations, thus avoiding injection of new-domain bias. Q2: The router is not updated without constraints. The gate uses token features and domain embeddings for content- and domain-aware routing, while Eq. 6 encourages domain-adaptive expert specialization. If router updates disrupt old-domain features, relational distillation penalizes the resulting relation-matrix changes, thereby constraining routing drift and preserving learned cross-domain relations. Q3: The P spatial tokens are randomly sampled from intermediate feature following ASP, with P=256.Each token corresponds to a local receptive field and encodes local morphology, e.g., tissue texture, cell distribution, and nuclear staining patterns (Eq.3), so we call them spatial descriptors. They are projected into a shared space (Eq.3) and processed by MoE (Eqs.4&5) for relation computation. Although the shared projection does not explicitly disentangle morphology and appearance, ASP’s HE-IHC spatial contrastive learning encourages spatial tokens to capture shared local structural features via shared projection, as HE and multi-IHC share morphology but differ in marker expression. The MoE further introduces domain-aware representations for comparable yet domain-specific relations. Q4: We agree that new-old relations are valuable. But this work focuses on preserving stable relations among stable old domains to mitigate forgetting, while the new domain is still being optimized and learned via paired supervision and generation loss. Tabs. 3&4 validate our current design, and we will explore new-old relations. Q5: Continual multi-domain virtual staining remains largely unexplored, with no directly comparable prior work. Our baselines cover representative CL mechanisms, including regularization, distillation, and replay, and are integrated into ASP (MICCAI 2023) for fair comparison. Results show generic CL methods are insufficient for this task. Q6: Initial denotes Seq-FT performance right after each domain is learned. For last domain Ki67, it equals Seq-FT and is not repeated.

R2 Q1: All CL baselines use ASP, with only different CL strategies integrated. Q2&Q3: Please refer to R1,Q3.Q4: MoE experts are not pretrained networks, but learnable MLP experts trained end-to-end with total loss. The gate is also a learnable MLP, taking token features and domain embeddings as input to generate expert weights via softmax. The final token representation is the weighted sum of expert outputs followed by L2 normalization. Q5&Q6: We follow PGVMS (TMI 2026) and use the same five metrics. PGVMS notes that PSNR/SSIM mainly measure pixel-level similarity and may be unreliable under HE-IHC spatial misalignment. PGVMS also adopts CONCH-FID, which measures distribution-level similarity using CONCH features that encode pathology-aware semantics, as CONCH is a pathology foundation model trained via pathology image-text contrastive learning. Thus, although our method may slightly trade off pixel-level accuracy, its better FID/CONCH-FID suggests improved pathology-aware staining quality.

R3 Q1: Please refer to R1,Q3.Q2: Memory cost: 5GB for 3 domains and 7GB for 4 domains. The main memory overhead comes from generating and encoding more old-domain images, which grows approximately linearly with the number of domains. Additional relation computation is performed sequentially over domain pairs, mainly increasing training time rather than peak memory. Thus, 10 domains would require about 20GB. Four-domain results show that our method can better mitigate forgetting. Larger-scale validation is future work due to costly paired multi-IHC data. Q3&Q4: We will fix the “L” and clarify “G” as gate net.



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 manuscript introduces a continual multi-domain virtual staining method for immunohistochemistry (IHC) images. It uses a mixture-of-experts (MoE)-based domain-aware feature modeling strategy to extract structured and disentangled representations for different domains. Then, it adopts a relational distillation mechanism to maintain cross-domain structural consistency across biomarkers. The method is evaluated on the MIST dataset, and the experiments show that it produces better virtual staining performance than several other relevant approaches. However, the reviewers have raised several significant concerns as follows:

    1.The algorithm design of the multi-domain relational distillation may be unstable. When a new domain is significantly different from previous ones, the estimated correlation matrix may be unreliable and the distillation could introduce significant bias.

    2.Updating the router may cause forgetting, which can negatively affect how the model understands cross-domain relationships. In addition, it may be not sufficient to merely maintain previously learned cross-domain relationships for fully addressing forgetting.

    3.The presentation of the method needs to be significantly improved (Reviewers #1 and #2). For instance, how to emphasize morphology-related structures in a unified latent space? What are the different experts and how are they trained? In addition, the gating network is not clearly discussed.

    4.The details of experimental setup are missing. For instance, do all continual learning-based baselines use ASP as the backbone? How the P spatial tokens are selected?

    5.The experiments are not inadequate. The study does not compare with recent state-of-the-art (SOTA) algorithms or models. Thus, it is difficult to evaluate the technical contributions of the study.

    6.The evaluation metrics used do not measure semantic similarities, and the ConchFID metric captures distributional similarities.

    7.In several cases (Tables 1 and 2), previous approaches outperform the proposed method in a few metrics, and this is not explained in the experiments. In addition, no computation cost is reported.

    Please consider addressing the reviewers’ concerns in the 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.

    This paper presents a continual multi-domain virtual staining method for immunohistochemistry (IHC) images. Specifically, it uses a mixture-of-experts (MoE)-based domain-aware feature modeling strategy to extract structured and disentangled representations for different domains. Then, it adopts a relational distillation mechanism to maintain cross-domain structural consistency across biomarkers. This paper studies an under-explored problem, i.e., continual multi-domain virtual staining, which may be an interesting research direction. In addition, the method produces promising experimental results and outperforms several other relevant approaches. The rebuttal addresses most of reviewer’s concerns, including clarification of algorithm design, experimental setup details, and computational cost. Thus, the manuscript is recommended for acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers found the paper timely and clinically relevant as the first work exploring continual learning for multi-domain virtual IHC staining. Strengths included the novel task formulation, the MoE-based domain-aware representation learning, and the relation-preserving distillation strategy. After rebuttal, two reviewers moved to accept and felt that most technical and clarity concerns were sufficiently addressed, while one reviewer maintained rejection due to remaining concerns about the robustness of the relational distillation mechanism and router forgetting. Overall, despite some remaining limitations regarding scalability, evaluation metrics, and theoretical justification, the paper was viewed as a promising and practically valuable contribution with convincing empirical performance.



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.

    This paper explores an interesting topic in multi-domain, continual learning-based virtual staining. The authors have adequately addressed the majority of the concerns, although some divergence of opinion remains among the reviewers.



back to top