List of Papers Browse by Subject Areas Author List
Abstract
Recently, virtual stain generation has emerged as a promising tool in histopathology, with the objective of enabling rapid and cost-effective staining procedures. While current staining protocols commonly rely on hematoxylin (H) and eosin (E), additional stains—such as saffron (S)—have demonstrated clinical value in pathological practice in some countries. In this paper, we investigate generative models to produce HES-stained images from standard HE-stained samples. The proposed method operates in the concentration domain, leveraging a stain separation algorithm to first extract concentration maps from the input HE images. A conditional Flow Matching framework is then developed to digitally generate the saffron stain, conditioned on the hematoxylin and eosin concentration maps. Finally, the RGB HES digital image is reconstructed according to the Beer–Lambert law. Extensive experiments on private and benchmark multi-tissue datasets demonstrate that the proposed approach produces high-fidelity HES images with strong visual realism. The source code as well as the subjective evaluation tool are available at: https://github.com/MedSalimBenOmrane/CFM-HES.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5285_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/5285_supp.zip
Link to the Code Repository
https://github.com/MedSalimBenOmrane/CFM-HES
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BenMoh_Making_MICCAI2026,
author = { Ben Omrane, Mohamed Salim AND Martin, Ségolène AND Kaaniche, Mounir AND Laurent-Bellue, Astrid AND Miled Souid, Wided AND Guettier, Catherine AND Pesquet, Jean-Christophe},
title = { { Making HE Histopathological Images More Colorful by Conditional Flow Matching } },
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
1.This paper proposes a control flow matching network to predict concentration maps. 2.The paper proposes an entire pipeline that first utilizes a state-of-the-art network for stain deconvolution, then employs the network proposed in the first point to predict the Saffron concentration map, and finally synthesizes the HES stained image. 3.This paper addresses the high cost of HES staining by using a generative model to obtain high-fidelity HES slides from HE-stained sections.
- 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 flow matching method proposed in this paper is sufficiently novel in the field of stain normalization. 2.The proposed workflow yields HES images with excellent visual performance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The comparative experiments in this paper do not fully demonstrate the advantages of CFM. 2.It is difficult to determine whether the superior performance is attributable to the CFM or state-of-the-art stain separation models.
- 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.
- 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
1.It would be beneficial to replace the entire CFM framework with a diffusion model for a more direct comparison. 2.It would be beneficial to conduct relevant ablation studies to demonstrate the effectiveness/contribution of using CFM for achieving the results presented in the paper.
- 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?
Overall, this paper demonstrates novelty by leveraging flow matching for stain normalization, achieving results comparable if not surpassing sota methods. However, based on the paper, it is unclear whether the performance improvement stems from the use of CFM or from the separation model. Therefore, I recommend a score of 4.
- 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.
After reading the PredS paper, we found that the data from comparative experiments can indeed be used to demonstrate the contribution of CFM to the final results. However, regarding the diffusion model, I think that considering only the improvement in speed is not sufficient to show that the diffusion model performs worse than CFM. In my view, an important comparative experiment is still missing.
Review #2
- Please describe the contribution of the paper
The primary contribution of the paper is the use of flow matching as a generative approach to synthesize Saffron staining from the hematoxylin and eosin components of an image. The novelty lies in leveraging stain deconvolution techniques to separate these components, enabling the model to more effectively generate the Saffron stain.
- 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 Major Strength of the paper are:
1.Using flow matching is relevant and currently popular in generative literature.
2.Manual evaluation by multiple pathologist to evaluate quality of the generated images.
- 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.
There are some major weaknesses in the paper:
1.Limited Methodological Novelty. There are already existing approaches that leverage stain separation techniques to generate missing stains, including cases similar to Saffron stain synthesis. Furthermore, the use of flow matching largely appears to be applied in an off-the-shelf manner with only minor modifications, which does not constitute a significant methodological contribution. As a result, the overall novelty of the work is limited, and the paper lacks a strong sense of innovation.
2.If the claimed novelty is that stain separation improves the quality of generated outputs, it would be more convincing to evaluate this explicitly by training other baseline methods both with and without stain separation. Such an ablation-style comparison would more directly demonstrate the contribution of the proposed idea.
Additionally, the choice of flow matching as the generative framework is not sufficiently justified. It would be important to clarify why flow matching is preferred over alternatives such as diffusion models, which are known to produce high-quality results in similar tasks. Is there evidence suggesting that flow matching performs better on smaller datasets or offers other practical advantages? Providing such justification, along with relevant comparisons, would strengthen the paper.
3.While the inclusion of manual evaluation is a valuable addition and the results are compelling, there is a notable limitation. The absence of comparisons with CycleGAN is concerning, as recent studies have shown that CycleGAN-based models can effectively reconstruct morphological and visual characteristics of stains—particularly when properly trained and well-converged. Therefore, the manual evaluation presented here feels incomplete without such a relevant baseline for comparison.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper exhibits limited methodological novelty, and the chosen comparisons are not sufficiently aligned with the proposed approach. As such, it appears more suitable for a workshop venue rather than a full conference. That said, the inclusion of manual evaluation is a strong aspect of the work and adds meaningful value.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal partially addresses some of my concerns; however, important issues still remain unresolved:
1.The authors justify the use of Flow-based models primarily on the basis that they require fewer inference steps. However, this argument is not entirely convincing, since diffusion models can also achieve efficient inference through accelerated sampling methods such as DDIM. While reducing the number of diffusion steps may introduce some degradation in generation quality, the rebuttal does not provide direct experimental evidence showing that the proposed Flow model achieves a superior efficiency–performance trade-off compared to optimized diffusion baselines.
A more convincing justification would require controlled comparisons between Flow and diffusion models under comparable inference budgets (e.g., equal latency or equal sampling steps). In the absence of such experiments, it remains unclear whether the observed advantage stems from the Flow formulation itself or simply from the selected inference configuration.
2.I acknowledge the authors’ clarification regarding the novelty of applying generative models to the H&E-to-HES staining task. However, the dataset used in this work has not been publicly released. As a result, the paper misses the opportunity to make a substantially stronger contribution by simultaneously introducing a benchmark dataset that could improve reproducibility and facilitate future research on this task.
Since the dataset is not publicly released, the broader impact and long-term utility of the work are somewhat limited, making it difficult for me to substantially increase my evaluation score despite the novelty of the application domain.
3.That said, I believe the released annotation application and accompanying code could still provide practical value to the community. Such a tool may be useful for other researchers working on annotation, qualitative assessment, and evaluation in virtual staining or related computational pathology tasks.
Overall, while the rebuttal improves my assessment of the work and could justify increasing the score, I am still not fully convinced by the methodological contribution of the paper. For me, the paper remains borderline accept at best; that said, I would not strongly oppose its acceptance or rejection. There is no borderline option otherwise I would have chosen that.
Review #3
- Please describe the contribution of the paper
This paper addresses the problem of virtual stain generation by synthesizing HES-stained histopathology images from standard HE images. The proposed method first performs stain separation to extract hematoxylin and eosin concentration maps from HE images. A conditional Flow Matching (CFM) framework is then used to generate the missing saffron concentration map conditioned on these observed stain components. Finally, the RGB HES image is reconstructed using the Beer–Lambert law. The method is evaluated on a private paired HE/HES liver dataset and qualitatively validated on a public multi-tissue dataset. Experimental results demonstrate improved reconstruction performance compared to several image translation baselines, and expert evaluation by pathologists suggests that the generated HES images are visually realistic.
Overall, the paper proposes a technically sound approach for virtual staining and explores a problem that is less frequently studied than traditional stain normalization.
- 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.Well-designed and physically grounded methodology
The proposed framework combines stain separation, conditional generative modeling, and Beer–Lambert reconstruction. By generating only the missing saffron concentration while preserving the observed hematoxylin and eosin components, the method is more structured than generic RGB-to-RGB image translation approaches and may reduce the risk of morphological artifacts.
2.Novel application of conditional flow matching
The use of conditional flow matching for virtual stain generation is technically interesting. Modeling the conditional distribution of saffron concentration maps given H/E concentrations provides a flexible formulation for this generative problem.
3.Less explored and potentially useful problem formulation
Compared to the large body of work on stain normalization, virtual stain generation from HE to HES addresses a different and relatively less explored task. This makes the paper more interesting from a methodological perspective and broadens the scope of computational pathology research.
4.Paired data and slide-level data split
The use of paired HE/HES WSIs and slide-level splitting reduces the risk of patch-level leakage and provides a more reliable experimental protocol.
5.Combination of quantitative and expert-based evaluation
The study reports both reconstruction metrics and pathologist-based subjective evaluation, which provides complementary evidence for the quality of the generated images.
- 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.
Weaknesses
1.Limited dataset size
The private paired dataset contains only 25 WSIs (with 2 used for testing). Although many patches are extracted, the limited number of independent slides restricts the strength of the conclusions and the assessment of generalization.
2.Lack of downstream task evaluation
The evaluation focuses primarily on reconstruction metrics and perceptual quality. While these metrics capture visual fidelity, it remains unclear whether the generated HES images improve or preserve performance in downstream pathology tasks such as classification, segmentation, or detection.
3.Limited analysis of cross-organ performance
The expert evaluation indicates weaker results for lung tissue compared to liver and breast samples. A deeper analysis of this behavior would strengthen the paper.
- 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.
- 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
Constructive Suggestions Include downstream pathology tasks to demonstrate the practical value of the generated HES images. Provide additional analysis of cases where saffron prediction is less accurate. Discuss more explicitly the potential clinical applications of virtual HES staining.
- 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?
This paper presents a well-designed and technically grounded approach for virtual stain generation from HE to HES images using conditional flow matching in the stain concentration domain. Compared with many recent works that revisit stain normalization, this paper addresses a less explored and more interesting problem formulation. The proposed pipeline—predicting only the missing saffron concentration while preserving hematoxylin and eosin components and reconstructing images through the Beer–Lambert model—is elegant and physically motivated.
While the experimental validation is somewhat limited by the size of the paired dataset and the absence of downstream task evaluation, the methodological contribution is clear and the results are promising.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I like the methodology from the beginning, only have concerns for the evaluation part, and the rebuttal is also nicely written and has addressed my concern.
Author Feedback
We sincerely thank the three reviewers and ACs for their valuable comments.Below, we address the main concerns. 1-Choice of Flow Matching over Diffusion Models (R1, R2, AC) Diffusion models suffer from two main limitations: they generally rely on Gaussian distributions, and there is a trade-off in selecting the time interval [0,T]. FM learns a constant-velocity field along the straight-line path, resulting in more stable training and much faster inference. These advantages, well established in the literature [16, 18], motivate our choice. Our CFM model generates high-quality saffron maps in only 50 ODE steps, while a standard DDPM requires about 1000 time steps. This speed-up is critical for whole-slide deployment. 2-Experimental Validation: Analysis of Cross-Organ Performance and Dataset Size (R3, AC) The lower performance on lung tissue is due to a domain shift: the model is trained only on liver WSIs, while lung tissue has a very different architecture (alveolar structure, low cellular density). This is acknowledged in the paper and motivates future work on cross-organ generalization. The slide-level split (21/2/2 WSIs) prevents any patch-level leakage, and 30,000 aligned HE/HES patches is a large dataset given the difficulty of obtaining registered slides. Generalization is also assessed on the public WSSB dataset. Subjective scoring by 5 pathologists on 84 images already provides strong clinical validation, and downstream task evaluation is left for future work. 3-Isolation of CFM’s Contribution from Stain Separation (R1, R2) The ablation requested by the reviewers is already reported in Tab. 1.PredS [5] follows the same pipeline as ours (stain separation, concentration domain processing, Beer-Lambert reconstruction) but uses a deterministic UNet instead of a conditional generative model. Our model improves PSNR from 31.34 to 32.06 dB, LPIPS from 0.047 to 0.030, and PieAPP from 1.06 to 0.66.Since the only architectural difference is the generative model, this comparison directly isolates the contribution of CFM. The improvement comes from modeling the full conditional distribution of saffron given H and E, rather than producing a single point estimate. 4-Concentration Domain versus RGB (R1) Working in the concentration domain is a physically motivated choice. An RGB model must reconstruct three channels at once, which can introduce artifacts on H and E even though they are already known. Our framework extracts H and E concentrations and keeps them fixed, while CFM generates only the missing saffron channel (Eq. 11). The task then reduces to single-channel conditional estimation, which is easier to optimize, and the observed stains are kept unchanged. This is an important property for clinical interpretation and acceptance of the method by clinicians. 5-Methodological Novelty (R2) We respectfully disagree that our method overlaps with prior work, and we would like to emphasize that stain separation has been widely used in the literature for stain normalization tasks, but not for virtual staining (except PredS [5]). Let us recall that RGB-based methods (CycleGAN, Pix2Pix, Pyramid Pix2Pix) perform holistic image translation without any physical grounding. PredS [5] operates in the concentration domain but relies on deterministic regression. To the best of our knowledge, no prior work formulates HE to HES translation as a conditional generative problem in the concentration domain, generating only the missing saffron stain while keeping H and E unchanged. This formulation is the main contribution of our work. 6-CycleGAN in Expert Evaluation (R2) Limiting the subjective evaluation to the top-performing methods is standard practice and was necessary to keep the workload manageable for the five pathologists. CycleGAN was the weakest baseline, with PSNR 4.41 dB below CFM. Evaluating all six methods would have tripled the number of images per pathologist (from 84 to 252) without adding useful information for the main comparison.
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 paper addresses an important question in digital pathology and has a technically sounds methodology. Reviewer concerns are around the limited experimental validation. Reviewers also raised the concern that the use of flow matching needs to be justified in more detail, e.g. does the use of flow matching consitute a significant methodological contribution and does it offer significant advantages.
- 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 work presents a creative virtual staining formulation that maps computer vision generation directly to physical pathology principles. Rather than relying on generic, artifact-prone RGB-to-RGB image translation, it maps only the single missing stain concentration map while preserving the underlying tissue architecture. The author’s rebuttal successfully demonstrated that their generative method significantly outperforms deterministic alternatives under identical pre-processing settings. The paper provides clear practical value and a sound methodology. Acceptance is recommended.
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 reviews remain somewhat mixed after revision. However, the majority of the reviewers view the revision positively and consider that the authors have adequately addressed the main concerns. Taking the overall discussion into account, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper applies Conditional Flow Matching for staining enrichment, which is an interesting direction. However, after reading the paper, review, and rebuttal, I remain concerned about the overall rigor and reliability of the evaluation. I agree with the reviewers that several important issues remain insufficiently addressed. One reviewer’s comment appear closer to borderline acceptance given unsolved concerns.
Overall, the submission ranks significantly below the 50th percentile of the rebuttal-stage papers in my batch. Given the number of unresolved concerns and the limited evidence provided to strengthen confidence in the methodology and evaluation, I am unable to support a positive decision