List of Papers Browse by Subject Areas Author List
Abstract
IgG4 immunohistochemical (IHC) staining plays a critical role in pathological diagnosis, where the quantitative assessment of IgG4-positive plasma cells—specifically their density and spatial distribution—serves as a pivotal pathological criterion for evaluating conditions like IgG4-related disease (IgG4-RD). H&E-to-IHC virtual staining offers a promising low-cost alternative to conventional immunostaining. However, most existing methods focus on global visual realism or continuous tissue textures (e.g., tumor regions), failing to capture the discrete and sparse diagnostic cell patterns required for IgG4 assessment. Moreover, real-world IHC images often contain non-specific staining (NSS), introducing noisy supervision that can be erroneously learned by generative models and compromise pathological consistency. Most existing virtual staining methods assume clean supervision or depend on strict data filtering, rather than explicitly accounting for NSS. To address these challenges, we propose IgG4-QPC, a virtual staining method that captures discrete IgG4-positive cell patterns from H&E images to achieve quantitative pathological consistency. Specifically, we introduce a Segmentation-Guided Instance-Level Alignment (SGIA) module, utilizing a segmentation prior to align cell-level features and enforce spatial density consistency, thereby enhancing sensitivity to diagnostically relevant plasma cells. Furthermore, we design a Noise-Aware Adversarial Learning strategy that identifies NSS artifacts and employs spatially-adaptive mask weighting to suppress background noise during optimization. Experimental results on paired H&E-IgG4 datasets demonstrate that our method significantly outperforms state-of-the-art approaches in image quality, cell density consistency, and spatial correlation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2533_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LuoMen_IgG4QPC_MICCAI2026,
author = { Luo, Mengyue AND Wu, Dawen AND Ji, Yuxuan AND Tian, Mingxuan AND Zhang, Haixian},
title = { { IgG4-QPC: Quantitative Pathology-Consistent Virtual Staining from H&E to IgG4 } },
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 IgG4-QPC, a virtual staining framework that translates H&E images to IgG4 IHC staining by explicitly preserving the discrete and sparse distribution of diagnostically relevant IgG4-positive plasma cells. The method introduces a Segmentation-Guided Instance-Level Alignment (SGIA) module and a Noise-Aware Adversarial Learning strategy to achieve quantitative pathological consistency under noisy supervision from non-specific staining artifacts.
- 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 well-defined and practically important challenge. 2.The proposed method is highly tailored to the IgG4 assessment task, and both components (SGIA and Noise-Aware Adversarial Learning) are directly motivated by the domain-specific challenges of sparse cell modeling and non-specific staining noise. 3.The paper is well written and easy to follow. The motivation, method design, and experimental results are logically organized, and the figures effectively illustrate the key ideas. 4.The use of a pre-trained, frozen segmentation network as a semantic prior to inject cell-level supervision is an elegant and practical design choice that avoids end-to-end training instability.
- 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 SGIA module relies critically on a pre-trained cell segmentation network S_G. However, the paper provides only a brief mention of this in the implementation details without elaborating on the data scale, annotation protocol, or generalizability of this segmentation model. 2.The identification of non-specific staining (NSS) artifacts via color deconvolution and segmentation priors is a reasonable and well-motivated strategy. However, the paper does not provide any qualitative or quantitative evaluation of the NSS detection pipeline itself. For example, visual examples of detected NSS masks, or comparison against manually annotated artifact regions. 3.All experiments are conducted on a single in-house dataset of orbital mass biopsies scanned at 40x magnification. While the dataset construction is carefully described, the generalizability of the proposed method to other tissue types, IgG4-related disease manifestations, or different scanning protocols remains unclear.
Minor: 1.Several instances of missing spaces after period signs are present throughout the manuscript. 2.Some sentences contain missing spaces between words and citations/parentheses.
- 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 paper presents a well-motivated, task-specific solution to an underexplored problem in computational pathology. The method design is clear and the quantitative results demonstrate meaningful improvements in pathology-oriented metrics. The primary concerns are the limited transparency in segmentation model training, the absence of evaluation for the NSS identification module, and the narrow evaluation scope restricted to a single dataset. Addressing these points would considerably strengthen the paper.
- 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 #2
- Please describe the contribution of the paper
This paper addresses H&E-to-IgG4 virtual staining for the quantitative assessment of IgG4-positive plasma cells, with a focus on preserving diagnostically relevant sparse and discrete cell patterns rather than only global visual realism. The authors highlight the challenge of non-specific staining (NSS) in real-world IHC images and propose IgG4-QPC, which combines a Segmentation-Guided Instance-Level Alignment (SGIA) module for enforcing cell-level feature and density consistency, and a Noise-Aware Adversarial Learning strategy for suppressing NSS artifacts during training. In addition, the authors construct a paired H&E-IgG4 dataset and demonstrate that the proposed method achieves improved performance over existing virtual staining approaches in both image quality and pathology-oriented quantitative metrics.
- 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 is well motivated, clearly identifying that IgG4 assessment requires instance-level quantitative consistency, which is not adequately addressed by prior virtual staining methods.
2.The problem of NSS as noisy supervision is practically relevant, and the paper explains this issue clearly with convincing motivation.
3.The overall framework is conceptually clean, with SGIA and noise-aware adversarial learning well aligned with the problem formulation.
4.Experimental results show consistent improvements over prior methods across both image quality and pathology-oriented metrics, and ablation studies support the contribution 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.
1.The method heavily relies on a pre-trained segmentation model to identify NSS and guide learning, but the reliability of this model is not sufficiently justified, raising concerns about error propagation.
2.The NSS handling strategy (e.g., thresholding, mask generation, and interpolation) lacks detailed analysis, and it is unclear how sensitive the method is to these design choices.
3.The evaluation depends on another frozen segmentation model (e.g., HoVerNet) for key pathology metrics, without sufficient validation of its suitability for this task.
4.Experiments are limited to a single dataset, making it difficult to assess generalization across different data distributions or clinical settings.
5.The framework requires additional components (e.g., segmentation prior), which may limit applicability to other datasets without extra annotation effort.
6.Some minor clarity and presentation issues remain (e.g., notation inconsistencies, insufficient explanation of intermediate variables).
- 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?
The paper addresses a meaningful and clinically relevant problem and proposes a reasonably well-motivated solution. The method design is coherent and shows promising results on the collected dataset. However, the core assumptions (especially NSS identification via segmentation) are not sufficiently validated, weakening the reliability of the proposed approach. The evaluation protocol and limited experimental scope further reduce confidence in the generalizability and robustness of the method. Overall, while the paper has potential, the current evidence is not strong enough to fully support the claims, leading to a current score.
- 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 #3
- Please describe the contribution of the paper
This paper addresses the specific challenge of H&E-to-IHC virtual staining for IgG4-related diseases, a clinical scenario where diagnosis heavily depends on the quantitative density and spatial distribution of discrete, sparse plasma cells rather than overall continuous tissue morphology. The authors’ main contribution is the development of the IgG4-QPC framework, which introduces two novel components designed to enhance instance-level pathological consistency. First, it proposes a Segmentation-Guided Instance-Level Alignment (SGIA) module that incorporates a pre-trained segmentation network as a semantic prior to enforce feature and density alignment between generated and real cell instances. Second, to tackle the practical issue of noisy supervision, the authors introduce a Noise-Aware Adversarial Learning strategy that identifies non-specific staining (NSS) artifacts and applies a spatially adaptive mask to suppress background noise during model optimization. The framework is empirically validated on a paired clinical dataset of orbital masses, demonstrating competitive image quality and improved quantitative pathological relevance compared to existing virtual staining baselines.
- 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.
This paper exhibits several notable strengths, particularly in its strong clinical motivation and highly targeted methodological design. First, the clinical problem is well-identified: the authors insightfully recognize that diagnosing IgG4-related diseases requires preserving discrete, sparse plasma cell instances rather than merely matching global tissue textures, addressing a significant limitation in standard image-to-image translation approaches. Second, the proposed IgG4-QPC framework introduces logically sound components to tackle this specific challenge, notably the Segmentation-Guided Instance-Level Alignment (SGIA) module that cleverly leverages semantic priors to enforce instance-level consistency, and a highly practical Noise-Aware Adversarial Learning strategy designed to mitigate the real-world issue of non-specific staining (NSS) artifacts. Finally, the evaluation methodology is comprehensive and well-aligned with the paper’s core claims; by incorporating pathology-specific quantitative metrics—such as cell density error (MAE) and spatial distribution correlation (Pearson-R)—alongside standard image quality assessments, the authors effectively demonstrate that their approach yields more reliable instance-level pathological consistency and cleaner backgrounds compared to conventional generative baselines.
- 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.
This paper addresses the clinically important problem of virtual staining from H&E to IgG4 for quantitative assessment of IgG4‑related disease, with a specific focus on generating sparse, discrete plasma cell instances and handling non‑specific staining (NSS) artifacts. The authors propose a Segmentation‑Guided Instance‑Level Alignment (SGIA) module that uses a pretrained segmentation network to align cell‑level features and enforce density consistency, and a Noise‑Aware Adversarial Learning strategy that identifies NSS artifacts and applies spatially adaptive weighting to suppress background noise. Experiments on a paired H&E‑IgG4 dataset of orbital mass WSIs show improved image quality and pathology‑relevant metrics over several GAN‑based baselines. However, several methodological and reporting concerns need to be addressed to strengthen the validity and reproducibility of the claims.
1.The noise‑aware cleaning pipeline relies on a frozen segmentation network to produce the cell probability map, which is then used to construct the NSS mask. If the segmentation network was trained on real IgG4 IHC images that already contain NSS artifacts, it may have learned to misinterpret some NSS patterns as positive cells, causing the cleaning algorithm to underestimate artifact regions and compromising the quality of the cleaned target image. The authors should clarify whether the segmentation training data included NSS examples, and ideally provide precision and recall of the segmentation network for distinguishing true plasma cells from NSS, either briefly in the main text or as supplementary material.
2.Key quantitative metrics in Tables 1 and 2 (PSNR, SSIM, MAE, PCC, KL) are reported only as point estimates without any indication of variability or uncertainty, making it difficult to judge whether the observed improvements are statistically reliable or could arise from random variation across test patches or slides. The only exception is KID which includes a standard deviation. The authors should compute and report standard deviations for these metrics (e.g., via bootstrap resampling or per‑slide statistics) and consider reporting p‑values from paired tests against baselines.
3.The paper does not provide a detailed description of how the segmentation network for positive plasma cells was trained and validated. It is mentioned that the network was pretrained using coarse annotations from Cellpose and fine‑tuned on a small set of pathologist‑annotated positive cells, but the size of the fine‑tuning set, the performance of the segmentation model (e.g., Dice, precision, recall on a held‑out test set), and whether it generalizes across different staining variations are not reported. Since the SGIA module heavily depends on the quality of this segmentation prior, the authors should report key segmentation metrics and discuss the potential impact of segmentation errors on the virtual staining results.
4.The comparison with existing virtual staining methods includes several GAN‑based approaches but does not include more recent diffusion‑based or transformer‑based virtual staining methods that have been published in the last two years. While the selected baselines are reasonable, the authors should justify the choice of baselines or at least discuss how the proposed method would compare to emerging architectures. Additionally, the paper does not include any evaluation by a pathologist (e.g., Turing test or diagnostic agreement study), which would be valuable to assess clinical utility beyond quantitative metrics.
5.The ablation study in Table 2 shows that removing the noise‑aware learning component (w/o NA) yields a model with MAE=3.2318 and PCC=0.3524, while the full model achieves MAE=3.4201 and PCC=0.3658.The trade‑off is that w/o NA has lower MAE but also lower PCC, and the differences are small. The paper does not discuss why the noise‑aware component slightly increases MAE while improving spatial correlation, nor does it provide statistical significance for these differences. The authors should provide a more nuanced interpretation of these results, including confidence intervals, to help readers understand the practical benefit of the noise‑aware learning strategy.
- 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.
(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?
My recommendation of a Weak Accept is primarily based on the paper’s strong clinical relevance and its targeted methodological approach to a nuanced problem in virtual staining, though some experimental constraints remain. The major factors influencing this score are:
The authors successfully identify a critical gap in existing H&E-to-IHC methods, which often prioritize global tissue textures over the discrete, sparse cell instances required for IgG4-related disease diagnosis. By shifting the focus toward quantitative pathological consistency through the SGIA module and noise-aware learning, the work provides a practical framework for generating diagnostically reliable virtual stains. Furthermore, the empirical results demonstrate a clear advantage in pathology-oriented metrics—such as cell density MAE and spatial correlation—over established generative baselines like CUT and CycleGAN. While the reliance on a single-organ dataset and a fixed segmentation prior for both training and evaluation suggests a need for further cross-domain validation, the current results offer a solid “proof-of-concept” that effectively suppresses false-positive artifacts caused by non-specific staining. Ultimately, the paper’s ability to bridge the gap between generative realism and clinical quantification outweighs its current limitations in dataset scale.
- 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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 specifically-designed virtual staining method to translate hematoxylin and eosin (H&E) images to IgG4 immunohistochemistry (IHC) data. Specifically, it designs a segmentation-guided instance-level alignment (SGIA) module that uses cell-level semantic priors to encourage cell-level feature alignment and density distribution consistency. Meanwhile, it adopts a noise-aware adversarial learning strategy to address the non-specific staining (NSS) artifacts for enhanced virtual staining. The method outperforms several generative adversarial network (GAN)-based approaches, and the ablation study demonstrates the effectiveness of each component. This work shows potential strong clinical impact.
However, the reviewers have pointed out some concerns, such as lack of implementation details of the pre-trained cell segmentation model, insufficient validation of the NSS identification module, and no standard deviation values reported for key evaluation metrics. Please consider addressing the reviewers’ concerns when preparing the final version.