Abstract

Virtual staining, which leverages generative artificial intelli- gence (AI) to produce immunohistochemistry (IHC)-stained tissue samples from hematoxylin and eosin (H&E)-stained images, has emerged as a cost-effective and accessible alternative to traditional IHC staining. Despite its potential, this approach faces three significant challenges: (1) the necessity of training a separate model for each tumor marker used in IHC staining, (2) the limited availability of large-scale datasets, and (3) the inherent diversity of staining patterns across different tissue types and markers. In this study, we address these challenges by introducing the Prompt-Driven Universal model for unpaired H&E-to-IHC Stain Translation (PD-UniST). Our approach incorporates two key innovations: (1) Structure-Cognizant Organization Prompt ModulE (SCOPE), which employs textual prompts to guide region-specific generation, and (2) Style-Prompt Unified Mapping ModulE (SPUME), which utilizes learnable prompts to capture task differences between various IHC stains and features a pathology-specific prompt-aware fusion layer for effective integration of visual features with task-specific prompts. Extensive experiments on two public datasets and one private dataset demonstrate that our method achieves state-of-the art performance across five different translation tasks, significantly improving both structural preservation and staining pattern accuracy. In clinical evaluation, we further validate the effectiveness of our method through pathologists’ assessment of both public and private datasets. The dataset and source code are available on anonymous GitHub at https://github.com/chujie-zhang/PD-UniST.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1617_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/chujie-zhang/PD-UniST.

Link to the Dataset(s)

SRRS datasete: https://drive.google.com/drive/folders/1rn9BgbaqwkijbvLSm3pmv8GMF67wed4r?usp=sharing

BibTex

@InProceedings{ZhaChu_PDUniST_MICCAI2025,
        author = { Zhang, Chujie and Xie, Yangyang and Li, Yinhao and Liang, Xiao and Lin, Lanfen and Chen, Yen-Wei},
        title = { { PD-UniST: Prompt-Driven Universal Model for Unpaired H&E-to-IHC Stain Translation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {462 -- 471}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel approach to virtual staining, which uses generative AI to produce immunohistochemistry (IHC)-stained tissue samples from hematoxylin and eosin (H&E)-stained images.

  • 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 SCOPE module effectively incorporates prior knowledge of subcellular localization patterns of different IHC markers by smartly using textual prompts (e.g., ‘cell nucleus’, ‘cell membrane’) to guide region-specific generation.

  • 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 authors do not provide an in-depth discussion in the text of the paper as to why certain methods are better at preserving structural similarity (SSIM) while others excel at generating more realistic and diverse images (FID).

    The experiment on the effectiveness of the stain transfer method is not sufficient. The paper only discusses how the proposed method improves IHC expression level prediction in Table 3. It would be expected to show more experiments on practical tasks (e.g. segmentation/classification).

  • 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.

  • 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/2025/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?

    Insufficient experiments, justification, and discussion.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes PD-UniST, a prompt-driven universal model for H&E-to-IHC stain translation, addressing three critical challenges: (1) marker-specific model training, (2) limited datasets, and (3) staining pattern diversity. Its core innovations include:SCOPE (Structure-Cognizant Organization Prompt ModulE) and SPUME (Style-Prompt Unified Mapping ModulE).

  • 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.

    Universal Modeling: Tackles marker/tissue diversity via prompt engineering (SCOPE/SPUME), reducing the need for separate marker-specific models. Clinical Validation: Pathologist evaluations on public/private datasets substantiate diagnostic utility—a rare but critical strength in virtual staining papers. Reproducibility: Experiments span three datasets (two public), enhancing credibility.

  • 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.Why are only “nucleus” and “membrane” prompts used? For markers like PD-L1 (which may show cytoplasmic staining), is this limiting? 2. Are prompts handcrafted or learned? If handcrafted, how were they validated for medical accuracy? 3.Does the cross-attention in SCOPE operate at the pixel, patch, or region level? This affects localization precision (critical for membrane vs. nucleus staining). 4.How does SCOPE handle ambiguous cases (e.g., mixed staining patterns or artifacts)? 5.The paper tests 5 markers (K=5)—how would scale to 20+ markers? Is there a bottleneck? 6. Add a scalability analysis (e.g., memory/runtime vs. K).

  • 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.

  • 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/2025/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 — could be accepted, 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 compelling advancement in virtual staining through its novel Prompt-Driven Universal model for H&E-to-IHC translation. The introduction of SCOPE (Structure-Cognizant Organization Prompt Module) and SPUME (Style-Prompt Unified Mapping Module) demonstrates significant technical innovation by effectively combining textual prompts with visual features through a biomedical BERT encoder and learnable task priors. The method’s ability to handle multiple tumor markers within a single framework addresses a critical limitation in the field, eliminating the need for marker-specific models. The paper’s strengths are particularly evident in its comprehensive clinical validation, including pathologist assessments on both public and private datasets - a rigorous approach that substantially strengthens its claims of clinical utility. The experiments show clear improvements over existing methods in both quantitative metrics and visual quality across five different translation tasks.

  • 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



Review #3

  • Please describe the contribution of the paper

    The authors propose a unified model that can perform virtual staining for multiple IHC tumor markers without requiring separate models for each task.

  • 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 integration of both textual and learnable task prompts is novel and well-motivated, offering a flexible and generalizable approach to virtual staining. +The authors conduct thorough quantitative, qualitative, and clinical evaluations, showing consistent improvements across multiple datasets and tasks. +PD-UniST outperforms state-of-the-art single-task and universal models in most metrics, especially in challenging settings like unpaired stain translation. +The paper includes detailed ablation analyses that clearly demonstrate the individual and combined benefits of the SCOPE and SPUME modules. +The involvement of expert pathologists in the evaluation enhances the practical validity 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.

    -I want to know, since this article has no supervised information, how can its precision and accuracy be guaranteed? -In Fig.3, there should be gaps between each row and column of images to facilitate observation.

  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Good motivation, good method, good writing

  • 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

Q1 (R1): Why only ‘nucleus’ and ‘membrane’ prompts: We focused on these main location features to test our method with a simple setup. Future work will add ‘cytoplasmic’ and other descriptors to capture more complex staining patterns Q2 (R1): Are prompts handcrafted or learned: We employ a hybrid system: textual prompts in SCOPE are manually crafted, while task prompts in SPUME are learnable. For the manually designed text prompts in the SCOPE module, we validate them through clinical experiments in Table 3. Q3 (R1): Does cross-attention in SCOPE operate at pixel, patch, or region level: SCOPE employs patch-level cross-attention, which balances precision and efficiency. This level preserves critical morphological distinctions between nuclei (circular internal structures) and membranes (ring-shaped peripheral structures). Pixel-level attention would be computationally prohibitive and noise-sensitive, while region-level attention would be too coarse to differentiate adjacent nuclear-membrane structures effectively. Q4 (R1): How does SCOPE handle ambiguous cases like mixed staining patterns or artifacts: SCOPE handles these challenges through its structure-aware cross-attention mechanism, which employs soft weighting when fusing BERT-processed text prompts with visual features. This assigns continuous weights rather than binary classifications, enabling representation of staining intensity gradients and positional uncertainty. The model was trained on complex clinical staining patterns and incorporates cycle-consistency and identity loss functions to maintain histological integrity in ambiguous regions, preventing over-interpretation of weak signals or artifacts. This combined approach enables robust handling of staining heterogeneity encountered in clinical practice. Q5 (R1): How would the model scale to 20+ markers: Although our initialized learnable task prior matrix is K×D, the matrix that participates in model training is 1×D, so regardless of whether there are 5 or 20 markers, the number of model parameters does not change. The main limiting factor is not computational capacity, but rather the ability to obtain diverse, high-quality training data. Q6 (R1): Add a scalability analysis: As we answered in the fifth question, for different values of K, memory usage and runtime do not change. Q7 (R2): In-depth discussion of the performance differences: ASP w/SCOPE gets better FID scores in single-task settings because its contrastive learning and focused attention creates more realistic textures when optimized for specific stains. This specialization better matches the patterns of individual markers. In contrast, our PD-UniST learns shared features across all six datasets, building stronger structural understanding but with slightly weaker modeling for individual datasets. Q8 (R2): The experiment on the effectiveness of the stain transfer method is not sufficient: Our task focuses more on testing IHC expression levels, so we prioritized showing pathologist clinical evaluations in Table 3. Due to space limitations, we didn’t include downstream task validation. We agree with this suggestion and will test the generated IHC images in downstream tasks in future work to better validate their clinical application potential. Q9 (R3): Since this article has no supervised information, how can its precision and accuracy be guaranteed: We ensure precision through complementary strategies: (1) Cycle-consistency loss enforces structural preservation; (2) PatchNCE contrastive loss maintains local tissue features; (3) SCOPE module incorporates domain knowledge via text prompts targeting biologically relevant structures; and (4) Most importantly, expert pathologist evaluation (Table 3) confirms the clinical accuracy of our generated IHC stains. Q10 (R3): In Fig.3, there should be gaps between each row and column of images to facilitate observation: Thank you for the reviewer’s comments, I will make revisions in the updated version.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top