Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Towards Reliable Stain Transfer: An Iterative Data-Model Co-Optimization Framework Based on Multimodal Expert-Guided Assessment

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An iterative expert-guided loop converts weakly paired H&E–IHC data into a state-of-the-art stain transfer model.

desk verdict Strong new framework for stain transfer, but the core selection mechanism needs an independent check before the SOTA claim fully lands. read the letter →

arxiv 2607.25393 v1 pith:4B7BYINL submitted 2026-07-28 cs.CV

Siyuan Xu , Yan Wang , Haofei Song , Lili Gao , Jiansheng Wang , Qing Zhang , Dan Huang , Boxiang Yun
show 2 more authors
Hongkai Xiong Qingli Li
This is my paper
classification cs.CV
keywords staintransferH&E-to-IHCimmunohistochemistrydata-modelco-optimizationvision-languagemodelmultimodalevaluationcomputationalpathologyweaklypairedimagetranslation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that H&E-to-IHC stain transfer can be made substantially more reliable by iteratively improving both the training data and the model, rather than by designing yet another generator architecture. It proposes DMCoStain, a three-stage pipeline: train a library of existing transfer models on weakly paired H&E and IHC images, generate pixel-aligned virtual IHC predictions from the training samples, then filter those predictions with a pathologist-style vision-language evaluator that keeps only samples judged clinically consistent with the real IHC images. The selected samples replace the weak pairs and the models are retrained, and the loop repeats. The paper claims this raises accuracy to state-of-the-art levels across six biomarkers from two tissue types, improves structural preservation, and yields an interpretable evaluation module that generalizes to biomarkers never seen during evaluator training.

What carries the argument

VLEGA combined with the answer-equality selection rule in Eq. (3) is the load-bearing mechanism: a generated IHC image is retained as a better-paired sample only if VLEGA's answers to four pathologist-designed clinical questions exactly match its answers for the corresponding real IHC image. This converts a weak, region-aligned supervision signal into a curated set of pixel-aligned pseudo-ground-truth pairs. The iterative pipeline (initial training, evaluation-and-selection, refined training, evaluation-and-selection, final training) is what turns those better pairs into better models, and a structural-preservation constraint (grayscale SSIM against the H&E image) prevents the selection proc

What would settle it

Take a held-out biomarker not used in VLEGA training, generate IHC from H&E with several transfer models, and ask VLEGA to accept or reject each candidate pair; then have a panel of pathologists independently flag clinically critical errors in the same pairs. If VLEGA's accepted set contains critical errors at the same rate as its rejected set, the Eq. (3) selection rule is not doing the work attributed to it. A cheaper complementary check is to ablate the loop with random selection of the same number of samples: if random selection reproduces the stage-wise gains, the co-optimization claim co

Watch

Extended reading notes

Core claim

The central discovery is a data-model co-optimization mechanism: instead of accepting pixel-unaligned H&E–IHC pairs as fixed supervision, DMCoStain generates on-train predictions, selects the ones that the Multimodal Expert-Guided Finer Selection (MEGFS) considers indistinguishable from real IHC on style fidelity, marker location, marker proportion and spatial position, and marker intensity, and retrains on those selected pairs. The selection is driven by VLEGA, a lightweight vision-language model trained on 147K instruction-answer pairs, whose answer-consistency rule (a generated image is kept only if VLEGA answers the same way for it and for the real IHC image) replaces coarse image-level

Load-bearing premise

The entire pipeline rests on trusting VLEGA's answer-consistency test—a generated image is kept only when a vision-language model trained on real IHC images answers four clinical questions the same way for the generated and real images—as an unbiased measure of pathological quality, including for biomarkers the model never saw. If that test is biased, the retraining loop will lock in the bias and the reported metric gains may not reflect true pathological fidelity.

Editorial extensions

If this is right

  • Any existing stain-transfer model can likely be plugged into the DMCoStain loop and improved without changing its architecture or loss design.
  • A library of imperfect weak-pair generators can bootstrap high-quality pixel-aligned supervision, reducing reliance on expensive aligned annotations.
  • MEGFS and VLEGA can function as a reusable automatic evaluation tool for future H&E-to-IHC models, including biomarkers absent from the evaluator's training set.
  • Gains propagate to downstream tasks: segmentation of molecular-positive regions on H&E images improves when the transferred IHC is used as an intermediate representation.
  • If the reported results hold, virtual IHC could lower the cost and turnaround of molecular staining in clinical pathology workflows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The selection rule has a self-referential flavor: VLEGA is trained on real IHC images from the same public datasets that serve as ground truth, so 'quality' is implicitly defined by those datasets' staining conventions; a new lab with a divergent staining protocol might see the loop over-reject legitimate variation.
  • A sharper test of clinical value would be whether DMCoStain improves agreement with pathologist biomarker scores (for example HER2 scoring) on a continuous scale, rather than only image-similarity and distribution metrics.
  • The same iterative select-and-retrain loop could transfer to other weakly paired image-translation problems, such as MRI-to-CT or fundus-to-OCT, whenever a reliable domain-expert evaluator is available.
  • Because VLEGA inference is fast (about 0.5 seconds per image), running more than the three tested rounds is cheap; whether additional iterations continue to yield gains is an empirical question the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DMCoStain, an iterative data-model co-optimization framework for H&E-to-IHC stain transfer under weakly paired supervision. Candidate stain transfer models are trained on weakly paired data, generate pixel-aligned predictions on the training set, and those predictions are filtered first by classical image-level metrics (CSS/PHV) and then by MEGFS, a VLM-based finer selection step. MEGFS retains a generated image only when VLEGA's answers to four pathologically motivated questions (style fidelity, marker location, proportion/spatial position, intensity) exactly match the answers for the corresponding real IHC image. The selected pseudo-pairs replace the original weak pairs and the process is repeated over three training stages. VLEGA is trained on ImmunoInstruction, a new 147,370-question VQA dataset built from MIST and HIT with expert refinement. The paper reports quantitative gains on MIST/HIT, a downstream segmentation ablation on a private PDAC dataset, blinded pathologist scoring, and VLEGA ablations.

Significance. If the claims hold, the paper makes a useful and timely contribution: a general data-model co-optimization loop that converts weak supervision into better-aligned pseudo-pairs, a new large-scale IPE instruction-following dataset, and a lightweight VLM evaluator for stain-transfer quality. The release of dataset/code and the use of downstream segmentation plus blinded pathologist evaluation are genuine strengths. However, the main SOTA claim is currently supported unevenly: Table 2 covers only four of the six claimed biomarkers, several gains over the strongest baselines are modest and carry no error bars, and the selection rule in Eq. (3) may encode VLEGA's training-distribution bias rather than verified pathological fidelity. These issues are addressable and do not invalidate the framework's potential, but they need to be fixed before the paper can be recommended for acceptance.

major comments (4)
  1. [Sec. 3.2, Eq. (3); Sec. 3.3; Table 5] Eq. (3) makes VLEGA's answer equality the sole fine-grained selection gate. VLEGA is trained on ImmunoInstruction, whose images are taken from the MIST/HIT training splits (Sec. 3.3), and the same public datasets supply the test sets for the final quantitative evaluation (Sec. 4.1). VLEGA is therefore a learned discriminator of the exact IHC distribution measured by CSS/PHV/FID/KID, so the refinement loop could improve these metrics by distribution matching even if VLEGA's judgments are not pathologically faithful. Table 5 shows VLEGA accuracy is only 80--91% internally and somewhat lower on held-out biomarkers, so the strict equality criterion is noisy. The PDAC segmentation and subjective evaluation partially mitigate this concern, but they do not isolate the selection criterion. Please report agreement between VLEGA and pathologists on generated candidates and ablate VLEGA selection a
  2. [Sec. 4.2, Table 2] The SOTA claim is not supported uniformly by the reported numbers. Table 2 shows results for only four biomarkers (ER, HER2, PAX5, CD3), although the abstract claims six; PR and Ki67 are deferred to the supplementary. On HITPAX5 and HITCD3, DMCoStain is worse than TDKStain on CSS (0.352 vs 0.427 and 0.400 vs 0.476) and on average PHV (0.375 vs 0.348 and 0.358 vs 0.323, where lower is better). DMCoStain wins on FID/KID, but the claim that it achieves SOTA “accuracy” across biomarkers requires a clear aggregation or a defensible metric hierarchy. Please present all six biomarkers and a summary statistic (e.g., average rank across metrics) or revise the claim.
  3. [Sec. 4.2.1, Table 2] No error bars, confidence intervals, or significance tests are reported for any quantitative metric. Many margins over the strongest baseline are small (e.g., ER CSS 0.155 vs 0.139; FID 32.1 vs 35.5), and single-run results are insufficient to establish a SOTA improvement, especially for a clinically oriented staining task. Please run at least 3 seeds for the central comparison, report mean and standard deviation, and use paired significance tests where appropriate.
  4. [Sec. 4.3.1, Eq. (1)] The paper says the numbers of Selected IHC Prediction per biomarker and stage are in the supplementary, but the main text does not report retention rates or threshold sensitivity. Since the CILPS thresholds in Eq. (1) are defined using averages of CSS/PHV over the candidate set, and since τ_struct, τ_k, θ, and model selection are free parameters, it is important to show how many pairs survive each evaluation-and-selection stage and whether the final results are stable to these thresholds. Without this, the data-refinement mechanism that is the paper's central contribution cannot be fully evaluated.
minor comments (5)
  1. [Abstract and Sec. 4.1] The abstract says “six biomarkers from two tissue types,” but Table 2 reports only four. Either add PR/Ki67 to the main table or explicitly state that they are in the supplementary and summarize their results.
  2. [Sec. 3.1, Eq. (1)] Please define ¯CSS and ¯PHV explicitly: are they averages over the candidate generated set, over the training set, or over the entire dataset?
  3. [Sec. 3.1, Eq. (2)] CILPSRT is used before being defined. Please clarify that it is the CILPS score evaluated after Refined Training.
  4. [Sec. 3.3] The citation for GPT-4o is [1], which is the GPT-4 technical report; please use an appropriate GPT-4o reference or clarify the version used.
  5. [Table 5] For the Random baseline, C3(2) (similarity score) is left empty; please explain why no random baseline is reported for that category.

Circularity Check

0 steps flagged · score 1.0 of 10

No construction-level circularity: the VLEGA-based selection gate is not the final evaluation metric; SOTA claims are anchored by held-out MIST/HIT test sets, private PDAC segmentation, and blinded pathologist scoring.

full rationale

The derivation chain is: weak pairs -> IT models -> CILPS -> MEGFS/VLEGA selection (Eq. 3) -> refined data -> RT -> FT -> evaluation. The only self-referential component is that VLEGA is trained on ImmunoInstruction, built from the training splits of the same public MIST/HIT datasets, and then used as the gate for pseudo-labels. This does not close the loop by construction, because the reported SOTA results are not scored by VLEGA or by Eq. (3). Table 2 uses held-out MIST/HIT test images with CSS/PHV/FID/KID; Table 3 and Fig. 6 use a private PDAC segmentation benchmark; Table 4 uses blinded pathologist scoring. None of these evaluation metrics is defined as VLEGA answer-equality, so the improvement is externally grounded rather than being the input to the derivation by construction. The paper's only self-citation is ATST-Net [36], used as one baseline/candidate model; it is not load-bearing because the framework's gains are shown across several independent models. There is no imported uniqueness theorem or ansatz from prior author work. A possible VLEGA bias is a validity concern, not a formal circularity, and per the review rules it is not treated as circular without a specific reduction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 4 invented entities

The framework depends on a small number of hand-set or data-dependent thresholds (τ_struct, mean-based CILPS filters, VLEGA evaluation tolerances) and on domain assumptions about VLEGA's judgment, the sufficiency of four evaluation dimensions, and the stability of self-training. These are not physical constants but design choices that could be varied; the paper does not fully audit their influence.

free parameters (4)
  • τ_struct (structural preservation threshold) = 0.6
    Hand-set threshold in Eq. (2b) to exclude structurally distorted models; no sensitivity analysis is reported.
  • CILPS thresholds (mean CSS, mean PHV) = adaptive per stage/biomarker
    In Eq. (1), candidates pass if CSS > mean and PHV < mean of the candidate distribution, a data-dependent filtering criterion that adapts to each model's output statistics.
  • VLEGA tolerance thresholds (τ_k, θ) = not specified in main text
    Proportion tolerance and GPT-similarity threshold for spatial-position evaluation are deferred to supplementary; they are free choices that affect which samples are selected.
  • Model selection for MEGFS per biomarker = best CILPS model
    Design choice to reduce compute; introduces a model-specific bias in the refined data.
assumptions (4)
  • domain assumption VLEGA answer equality is a valid proxy for IHC staining quality
    Central selection criterion in Eq. (3); assumes consistency of a VLM's answers across four categories indicates clinically acceptable staining.
  • domain assumption The four evaluation categories (style, location, proportion/position, intensity) capture the clinically relevant dimensions of staining quality
    Chosen by pathologists; not proven to be complete or sufficient, but reasonable.
  • domain assumption Selected on-train generated images provide better supervision than original weakly paired real images
    Central to the co-optimization loop; if the fakes are as noisy as the weak pairs, the loop won't help. The paper provides indirect evidence (improvements) but does not isolate this assumption.
  • domain assumption Models can be retrained on their own selected outputs without performance collapse
    Self-training stability is assumed; the paper shows improvements but no analysis of failure modes.
invented entities (4)
  • VLEGA independent evidence
    purpose: Vision-language model for IHC-positive-expression assessment used as the selection judge in MEGFS
    Paper claims release of code/weights; not independently verified, but a falsifiable handle is nominally provided.
  • MEGFS independent evidence
    purpose: Multimodal expert-guided finer selection strategy for filtering generated IHC images
    Algorithmic strategy; others can re-implement and test it on public data.
  • ImmunoInstruction independent evidence
    purpose: Large-scale IHC instruction-following dataset (150K VQA pairs) for training VLEGA
    Claimed to be released; gives a reusable resource.
  • DMCoStain independent evidence
    purpose: Overall iterative data-model co-optimization training framework
    Framework artifact; code claimed to be released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Reliable Stain Transfer: An Iterative Data-Model Co-Optimization Framework Based on Multimodal Expert-Guided Assessment." pith.science (2026). https://pith.science/paper/4B7BYINL

@misc{pith2026260725393,
  author       = {Pith},
  title        = {Pith review of: Towards Reliable Stain Transfer: An Iterative Data-Model Co-Optimization Framework Based on Multimodal Expert-Guided Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4B7BYINL}},
  note         = {Machine review of arXiv:2607.25393}
}
read the original abstract

Histopathological examination primarily relies on hematoxylin and eosin (H&E) and immunohistochemistry (IHC) staining. Although IHC provides critical molecular information, it is costly and requires specialized expertise. Stain transfer provides an efficient alternative by computationally generating IHC from H&E images, but remains challenged by unified and interpretable modeling for heterogeneous biomarkers under pixel-unaligned supervision. We propose DMCoStain, a novel Data-Model Co-optimization framework for Stain transfer. It iteratively co-refines training data and model capability, improving staining accuracy and interpretability in both pathological and structural consistency. To refine training data in a clinically meaningful manner, it incorporates the Multimodal Expert-Guided Finer Selection (MEGFS) strategy, built upon a pioneering IHC-positive-expression (IPE) vision-language model (VLM) that emulates pathologist reasoning. To support MEGFS, we construct ImmunoInstruction, the first large-scale IPE instruction-following dataset with 150K VQA samples. Extensive experiments on multiple tissues and biomarkers demonstrate that DMCoStain achieves state-of-the-art (SOTA) accuracy. This paradigm offers strong practical value, and MEGFS also functions as a specialized evaluation tool for future model development. Dataset, code, and more details are in https://github.com/SikangSHU/DMCoStain.

Figures

Figures reproduced from arXiv: 2607.25393 by the authors.

Figure 1
Figure 1. (a) Workflow for H&E and IHC image preparation; [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Classical evaluation fails to detect fine-grained local [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed DMCoStain. For training data, darker colors indicate a higher proportion of better-paired [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on MIST and HIT. Rows show ER, HER2, PAX5, and CD3 from top to bottom, at 1024 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Segmentation results on private PDACCK test data across stages. D = Dice, I = IoU. All images are 512×512. pathological consistency. In the red-boxed regions, both the posi￾tion and intensity of positive expression in the generated images align more closely with the re…
Figure 5
Figure 5. Figure 5: PHV trends of On-Test Prediction across stages. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 3 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.ArXiv Preprint ArXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Joseph Boyd, Irène Villa, Marie-Christine Mathieu, Eric Deutsch, Nikos Paragios, Maria Vakalopoulou, and Stergios Christodoulidis. 2022. Region-guided cyclegans for stain transfer in whole slide images. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. 356–365

    2022

  3. [3]

    Fuqiang Chen, Ranran Zhang, Boyun Zheng, Yiwen Sun, Jiahui He, and Wenjian Qin. 2024. Pathological semantics-preserving learning for H&E-to-IHC virtual staining. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention. 384–394

    2024

  4. [4]

    Ying Chen, Guoan Wang, Yuanfeng Ji, Yanjun Li, Jin Ye, Tianbin Li, Ming Hu, Rongshan Yu, Yu Qiao, and Junjun He. 2025. Slidechat: A large vision-language assistant for whole-slide pathology image understanding. InProceedings of the Computer Vision and Pattern Recognition Conference. 5134–5143

    2025

  5. [5]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://vicuna.lmsys.org

    2023

  6. [6]

    Tong Ding, Sophia J Wagner, Andrew H Song, Richard J Chen, Ming Y Lu, Andrew Zhang, Anurag J Vaidya, Guillaume Jaume, Muhammad Shaban, Ahrong Kim, et al. 2025. A multimodal whole-slide foundation model for pathology.Nature Medicine(2025), 1–13

    2025

  7. [7]

    Omar SM El Nahhas, Chiara ML Loeffler, Zunamys I Carrero, Marko van Treeck, Fiona R Kolbinger, Katherine J Hewitt, Hannah S Muti, Mara Graziani, Qinghe Zeng, Julien Calderaro, et al . 2024. Regression-based deep-learning predicts molecular biomarkers from pathology slides.Nature Communications15, 1 (2024), 1253

    2024

  8. [8]

    Xianchao Guan, Yifeng Wang, Ye Zhang, Zheng Zhang, and Yongbing Zhang. 2025. OT-StainNet: Optimal transport driven semantic matching for weakly paired h&e-to-ihc stain transfer. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 3194–3202

    2025

Show all 46 references
  1. [9]

    Xianchao Guan, Zheng Zhang, Yifeng Wang, Yueheng Li, and Yongbing Zhang

  2. [10]

    Yufang He, Zeyu Liu, Mingxin Qi, Shengwei Ding, Peng Zhang, Fan Song, Chen- bin Ma, Huijie Wu, Ruxin Cai, Youdan Feng, et al . 2024. PST-Diff: achieving high-consistency stain transfer by diffusion models with pathological and struc- tural constraints.IEEE Transactions on Medi...

    2024

  3. [11]

    Shengyi Hua, Fang Yan, Tianle Shen, Lei Ma, and Xiaofan Zhang. 2024. PathoDuet: Foundation models for pathological slide analysis of H&E and IHC stains.Medical Image Analysis97 (2024), 103289

    2024

  4. [12]

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-to- image translation with conditional adversarial networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1125–1134

    2017

  5. [13]

    Pascal Klöckner, José Teixeira, Diana Montezuma, João Fraga, Hugo M Horlings, Jaime S Cardoso, and Sara P Oliveira. 2025. H&E to IHC virtual staining methods in breast cancer: an overview and benchmarking.npj Digital Medicine8, 1 (2025), 384

    2025

  6. [14]

    Fangda Li, Zhiqiang Hu, Wen Chen, and Avinash Kak. 2023. Adaptive super- vised patchnce loss for learning H&E-to-IHC stain translation with inconsistent groundtruth image pairs. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. 632–641

    2023

  7. [15]

    Junjian Li, Hulin Kuang, Jin Liu, Hailin Yue, Mengshen He, and Jianxin Wang

  8. [16]

    Junjian Li, Jin Liu, Hulin Kuang, Hailin Yue, Mengshen He, and Jianxin Wang

  9. [17]

    Yuci Liang, Xinheng Lyu, Wenting Chen, Meidan Ding, Jipeng Zhang, Xiangjian He, Song Wu, Xiaohan Xing, Sen Yang, Xiyue Wang, et al. 2025. Wsi-llava: A multimodal large language model for whole slide image. InProceedings of the IEEE/CVF International Conference on Computer Visi...

    2025

  10. [18]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in Neural Information Processing Systems36 (2023), 34892–34916

    2023

  11. [19]

    InInternational Conference on Medical Image Com- puting and Computer-Assisted Intervention

    MiCo: Multiple Instance Learning with Context-Aware Clustering for Whole Slide Image Analysis. InInternational Conference on Medical Image Com- puting and Computer-Assisted Intervention. Springer, 376–385

  12. [20]

    Shengjie Liu, Chuang Zhu, Feng Xu, Xinyu Jia, Zhongyue Shi, and Mulan Jin. 2022. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix. InProceedings of the Computer Vision and Pattern Recognition Conference. 1815–1824

    2022

  13. [21]

    Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. 2024. A visual-language foundation model for computational pathology.Nature Medicine 30, 3 (2024), 863–874

    2024

  14. [22]

    Shuting Liu, Baochang Zhang, Yiqing Liu, Anjia Han, Huijuan Shi, Tian Guan, and Yonghong He. 2021. Unpaired stain transfer using pathology-consistent con- strained generative adversarial networks.IEEE Transactions on Medical Imaging 40, 8 (2021), 1977–1989

    2021

  15. [23]

    Marc Macenko, Marc Niethammer, James S Marron, David Borland, John T Woosley, Xiaojun Guan, Charles Schmitt, and Nancy E Thomas. 2009. A method for normalizing histology slides for quantitative analysis. In2009 IEEE Interna- tional Symposium on Biomedical Imaging: From Nano to...

    2009

  16. [24]

    Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. 2020. Contrastive learning for unpaired image-to-image translation. InEuropean Conference on Computer Vision. Springer, 319–345

    2020

  17. [25]

    Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al

  18. [26]

    Qiong Peng, Weiping Lin, Yihuang Hu, Ailisi Bao, Chenyu Lian, Weiwei Wei, Meng Yue, Jingxin Liu, Lequan Yu, and Liansheng Wang. 2024. Advancing H&E- to-IHC virtual staining with task-specific domain knowledge for HER2 scoring. InInternational Conference on Medical Image Comput...

    2024

  19. [27]

    Yue Peng, Bing Xiong, Fuqiang Chen, Deboch Eybo Abera, RanRan Zhang, Wan- ming Hu, Jing Cai, and Wenjian Qin. 2026. Usigan: Unbalanced self-information feature transport for weakly paired image ihc virtual staining.IEEE Transactions on Image Processing(2026)

    2026

  20. [28]

    Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. 2002. Color transfer between images.IEEE Computer Graphics and Applications21, 5 (2002), 34–41

    2002

  21. [29]

    Pushpak Pati, Sofia Karkampouna, Francesco Bonollo, Eva Compérat, Martina Radić, Martin Spahn, Adriano Martinelli, Martin Wartenberg, Marianna Kruithof- de Julio, and Marianna Rapsomaniki. 2024. Accelerating histopathology work- flows with generative AI-based virtually multipl...

    2024

  22. [30]

    Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Kr- ishna, and Linda Shapiro. 2024. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. InProceedings of the IEEE/CVF Conference on Computer Vision...

    2024

  23. [31]

    Gil Shamai, Amir Livne, António Polónia, Edmond Sabo, Alexandra Cretu, Gil Bar-Sela, and Ron Kimmel. 2022. Deep learning-based image analysis predicts PD-L1 status from H&E-stained histopathology images in breast cancer.Nature Communications13, 1 (2022), 6753

    2022

  24. [32]

    Yuxuan Sun, Yixuan Si, Chenglu Zhu, Xuan Gong, Kai Zhang, Pingyi Chen, Ye Zhang, Zhongyi Shui, Tao Lin, and Lin Yang. 2025. Cpath-omni: A unified multimodal foundation model for patch and whole slide image analysis in compu- tational pathology. InProceedings of the Computer Vi...

    2025

  25. [33]

    AC Ruifrok and DA Johnston. 2001. Quantification of histochemical staining by color deconvolution.Analytical and Quantitative Cytology and Histology23, 4 (2001), 291–299

    2001

  26. [34]

    Qwen Team et al. 2024. Qwen2 technical report.ArXiv Preprint ArXiv:2407.10671 2, 3 (2024)

    2024 arXiv

  27. [35]

    Jinxi Xiang, Xiyue Wang, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, et al. 2025. A vision–language foundation model for precision oncology.Nature638, 8051 (2025), 769–778

    2025

  28. [36]

    Siyuan Xu, Haofei Song, Yingjiao Deng, Jiansheng Wang, Yan Wang, and Qingli Li

  29. [37]

    Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. 2024. Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. InProceedings of the AAAI Conference on Artificial Intell...

    2024

  30. [38]

    Wenchuan Zhang, Jingru Guo, Hengzhe Zhang, Penghao Zhang, Jie Chen, Shuwan Zhang, Zhang Zhang, Yuhao Yi, and Hong Bu. 2026. Patho-AgenticRAG: Towards multimodal agentic retrieval-augmented generation for pathology VLMs via reinforcement learning. InProceedings of the AAAI Conf...

    2026

  31. [39]

    Wei Zhang, Tik Ho Hui, Pui Ying Tse, Fraser Hill, Condon Lau, and Xinyue Li. 2024. High-resolution medical image translation via patch alignment-based bidirectional contrastive learning. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. 178–188

    2024

  32. [40]

    Wenchuan Zhang, Penghao Zhang, Jingru Guo, Tao Cheng, Jie Chen, Shuwan Zhang, Zhang Zhang, Yuhao Yi, and Hong Bu. 2025. Patho-R1: A multi- modal reinforcement learning-based pathology expert reasoner.ArXiv Preprint ArXiv:2505.11404(2025)

    2025

  33. [41]

    InProceedings of the 34th Interna- tional Joint Conference on Artificial Intelligence

    Advancing stain transfer for multi-biomarkers: A human annotation-free method based on auxiliary task supervision. InProceedings of the 34th Interna- tional Joint Conference on Artificial Intelligence. 2116–2124. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Siyuan Xu et al

    2026

  34. [42]

    Zhe Xu, Ziyi Liu, Junlin Hou, Jiabo Ma, Cheng Jin, Yihui Wang, Zhixuan Chen, Zhengyu Zhang, Fuxiang Huang, Zhengrui Guo, et al. 2025. A versatile pathology co-pilot via reasoning enhanced multimodal large language model.ArXiv Preprint ArXiv:2507.17303(2025)

    2025 arXiv

  35. [46]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. InPro- ceedings of the IEEE International Conference on Computer Vision. 2223–2232

    2017

  36. [2024]

    A multimodal generative AI copilot for human pathology.Nature634, 8033 (2024), 466–473

    2024

  37. [2025]

    Supervised Information Mining from Weakly Paired Images for Breast IHC Virtual Staining.IEEE Transactions on Medical Imaging44, 5 (2025), 2120–2130

    2025

  38. [2026]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Universal-to-Specific: Dynamic Knowledge-Guided Multiple Instance Learning for Few-Shot Whole Slide Image Classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 26614–26623

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.