Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Harnessing Adversarial Distillation to Customise Debiased, Disease-Specific Pathology Foundation Models for Breast Cancer

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A compact breast-cancer-specific model can be distilled from billion-parameter pathology foundation models while suppressing site-related bias, and match them on external cohorts at 30x smaller size.

desk verdict A compact-distillation recipe that deserves a review, but the debiasing mechanism needs a direct site-probe measurement before the title claim is supportable. read the letter →

arxiv 2608.01356 v1 pith:XVLQQUVR submitted 2026-08-02 cs.CV

Zhiwei Chen , Yang Hu , Yuxiang Xiao , Yakun Ju , Tianyang Zhang , Yingxue Xu , Wei Li , Hao Chen
show 2 more authors
Jens Rittscher Kaixiang Yang
This is my paper
classification cs.CV
keywords pathologyfoundationmodelsknowledgedistillationadversarialdebiasingbreastcancerwhole-slideimagingself-supervisedlearningsiteconfoundingmultipleinstance
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

SmartStu is a recipe for turning big, general-purpose pathology foundation models into small, breast-cancer-specific encoders without inheriting the site-related bias the big models carry. The paper's core assertion is that by distilling from several frozen teachers into a lightweight student while simultaneously training that student to suppress features a learned 'noise model' uses to predict a slide's site of origin, a 1-24-million-parameter model can preserve, and at times improve, the balanced accuracy and AUC of 300-680-million-parameter teachers on external breast-cancer cohorts. This matters because deployment of foundation models in clinical pathology is currently gated by compute and by distribution shift; the result points to a controllable middle path: keep the useful morphology, drop the nuisance.

What carries the argument

Adversarial distillation against a learned noise model. A small auxiliary network is first trained to predict site labels on the multi-site distillation repository, optimising a supervised contrastive loss plus cross-entropy, and is then frozen. During student training, the student's features are projected and passed through a gradient reversal layer, and cosine similarity measures how well they align with the frozen noise model's site signature: $\mathcal{L}_{\mathrm{adv}} = 1 - \mathrm{CosSim}(z^{\mathrm{adv}}_i, h^{\mathrm{adv}}_i)$. Since the gradient is reversed, minimising this term makes the student representation worse at the nuisance-prediction task. The student also minimises mean-

What would settle it

Train a site classifier on frozen teacher features and on SmartStu student features over a held-out subset of the 40-site distillation repository, using the same site labels. If the classifier's accuracy on student features is not substantially lower than on teacher features, the adversarial loss is not actually suppressing site cues. Repeating this on any external cohort that carries centre metadata would test whether the debiasing transfers.

Watch

Extended reading notes

Core claim

SmartStu's claim is that a small, disease-specific pathology encoder can inherit the useful morphology of large foundation models without inheriting their site-specific bias. The method's discovery is that this can be done with adversarial distillation: a noise model is first trained to predict a tile's site of origin, then used as a counterexample while the student is trained; a gradient reversal layer forces the student's features to become less useful for site prediction while still matching teacher features. In their experiments, students with under 25 million parameters match or sometimes beat the frozen features of teachers in the 300M-680M range on three external breast cancer cohorts

Load-bearing premise

The load-bearing premise is that the site-related signatures the noise model learns on the 40-site distillation repository are the same kind of non-biological nuisance that causes failures on the external cohorts, and that removing them does not erase morphology needed for breast-cancer diagnosis; the paper does not measure either directly.

Editorial extensions

If this is right

  • A breast-cancer-specific tile encoder of 1-24 million parameters can be deployed where 300-680-million-parameter teachers cannot, while largely preserving external-cohort accuracy.
  • The adversarial component adds a consistent stepwise gain over plain distillation across backbones and cohorts, so the bias-suppression term, not just the extra teachers, is doing work.
  • Because the student is a standalone encoder, any standard multiple-instance-learning aggregator can consume its embeddings; the customisation happens once, at distillation time.
  • Changing the nuisance labels used to train the noise model should steer the student away from other non-biological signatures, such as stain, scanner, or tissue-processing artefacts, without retraining the teachers.
  • Very small students around one million parameters do not always benefit from the extra self-supervised loss, so the full recipe should be reserved for backbones with enough capacity.

Reading between the lines

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

  • The paper never directly tests whether student embeddings are less site-predictive than teacher embeddings; a site classifier trained on student features across the 40-site repository would provide that test and could be run without new data.
  • The noise model's nuisance axes come from a 40-site repository; whether those axes align with the nuisance variation in the external cohorts is untested, and an external cohort with site metadata would separate debiasing from generic regularisation.
  • The same framework can be applied to other disease types or to single-centre customisation by choosing the nuisance target to match the deployment setting; this is a direct extension the authors gesture at in the conclusion.
  • The observed performance-efficiency trade-off suggests the task-relevant information is far more compressible than the teacher parameter count implies; varying student capacity systematically would map where compression begins to lose diagnostic morphology.
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

3 major / 5 minor

Summary. The paper proposes SmartStu, a framework that distils knowledge from multiple large pathology foundation models (PFMs) into compact, breast-cancer-specific student encoders while attempting to suppress site-related confounding via an adversarial distillation loss. The student is trained on TCGA-BRCA tiles with a noise model pre-trained to predict site labels, and the objective combines multi-teacher MSE distillation, a gradient-reversal cosine-similarity loss against the frozen noise model, and a DINO-style self-supervised term with artefact injection. The method is evaluated on three external cohorts (Yale HER2, SLN-Breast, BRACS) across several small backbones, reporting balanced accuracy and AUC. The authors claim that SmartStu produces models over 30x smaller than general PFMs while preserving or sometimes improving downstream performance, and that the adversarial component actively suppresses site-related bias.

Significance. If the central claims hold, the paper would make a useful practical contribution: compact disease-specific encoders could replace billion-parameter PFMs in resource-constrained pathology workflows, and adversarial distillation could provide a principled way to remove site-specific nuisance signals during knowledge transfer. The evaluation design has clear strengths: the external cohorts are not used for training, hyperparameters appear fixed a priori, multiple backbones and teacher combinations are tested, and code is released. The reported improvements are consistent in direction across most backbones and cohorts. However, the load-bearing debiasing claim is not directly verified: the paper never shows that student features become less predictive of site, nor that the TCGA-BRCA site signatures are the nuisances present in the external cohorts. In addition, all quantitative claims are based on means over five runs without measures of variance or significance, so the magnitude and reliability of the reported gains are uncertain.

major comments (3)
  1. [§3.1, Tables 1 and 2] The adversarial loss is not a genuine adversarial procedure. The noise model T_adv is pre-trained and frozen, and the student feature z_s is first mapped by a learned projection head ψ_adv before the GRL and cosine-similarity loss. A frozen adversary cannot re-discover residual site-predictive directions that remain after the student updates, and the projection head can absorb the gradient-reversal pressure by discarding only the components needed to reduce similarity to T_adv's embedding. Consequently, the backbone representation z_s may still encode site information that a downstream linear or nonlinear classifier could exploit. To support the central 'debiasing' claim, the authors should directly measure site-predictiveness (e.g., train a site classifier on frozen student features before and after training, or evaluate clustering/alignment by site) and show that site information is ac
  2. [§2.2 and §3.1] All reported results are means of five runs with no standard deviations, confidence intervals, or significance tests, yet many of the claimed improvements are numerically small. For example, in Table 2, several Adv vs Itg differences are within 0.01 in bAcc or AUC (e.g., ResNet50 AdvS vs Adv on Yale bAcc: 0.8674 vs 0.8582 is larger, but TinyViT-5M Adv vs AdvS differ by only ~0.002 in AUC; TinyViT-21M AdvS vs Adv on Yale bAcc: 0.8974 vs 0.8974). Without uncertainty estimates, the reader cannot judge whether the stepwise improvements attributed to adversarial distillation and self-supervision are meaningful, especially for the smaller backbones. The authors should report per-run results or paired significance tests (e.g., across the five runs) for the key comparisons, at least for the ablation rows and the headline 'sometimes improving' claims.
  3. [§3.2, Figure 2] The debiasing mechanism rests on the assumption that site labels in TCGA-BRCA are a valid proxy for the non-biological nuisance signals that affect the external cohorts, and that suppressing these site cues does not remove diagnostically relevant morphology. This assumption is never tested. The paper provides no evidence that Yale HER2, SLN-Breast, or BRACS share the same site signatures as TCGA-BRCA (they likely have different scanning protocols, stain batches, and tissue types), nor whether the adversarial loss preferentially removes low-level edge/stain cues rather than subtle morphology. The consistent downstream gains could alternatively arise from a generic regularisation effect of the extra loss, rather than from specific removal of shared nuisance directions. The authors should add an analysis of what is being suppressed, for example by evaluating student features on external coh
minor comments (5)
  1. [Throughout] The metric name is inconsistent: 'balanced accuracy' is abbreviated as BACC in §3.1 and as bAcc in the tables and abstract. Please standardise.
  2. [§3.1] Only one train/test split is used for each external cohort (e.g., 0.8/0.2 for Yale-HER2 and SLN-Breast). Since the cohorts are small, results could be sensitive to the split. Reporting a few different splits or cross-validation would strengthen the conclusions.
  3. [Figure 2] The interpretability panels (C and D) are single-example qualitative visualisations. They are supportive but not evidence for the debiasing claim; consider quantifying attention overlap with ROIs across the test set.
  4. [§2.2] The phrase 'adversarial teacher' is used for the frozen noise model, but the model does not update adversarially during student training. This terminology may mislead readers; a short clarification is needed.
  5. [Abstract and §3.3] The claim 'over 30x smaller' is based on parameter counts, while the performance-efficiency trade-off in Fig. 2-B uses GFLOPs. These are different efficiency axes; the text should state which is used in the 30x claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: external-cohort evaluation is self-contained; minor self-citations are not load-bearing.

full rationale

The paper's central claim is supported by distillation from frozen teacher PFMs and evaluation on three external cohorts (Yale HER2, SLN-Breast, BRACS) that are not used during training. Hyperparameters (lr=1e-4, weight decay=0.05, lambda_adv=1e-4, lambda_ssl=1e-4) are fixed and reported a priori; no target metric is fitted on the test sets. The adversarial noise model is trained once on TCGA-BRCA site labels, frozen, and used as a fixed nuisance target; the student is trained to suppress alignment to that target via GRL. This is a training objective, not a renamed prediction or a self-defined equivalence: whether it actually removes site cues is an empirical question, and the paper's downstream gains are measured externally. The only self-citations are contextual (ref. [23] survey on PFM deployment, ref. [25] whole-slide foundation model) and do not carry the derivation of SmartStu or the evaluation. The fact that the student is trained on breast-cancer tiles makes the model 'breast-cancer-specific' by data construction, but the performance claim is not tautological because it is verified on unseen cohorts against frozen teacher features and ImageNet-initialized baselines. The untested transferability of TCGA-BRCA site signatures to external cohorts is a correctness/validity concern, not a circularity one.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability of semantic teachers, on the validity of site labels as a nuisance proxy, and on the transferability of the learned suppression to unseen cohorts. No unobserved physical entities are introduced.

free parameters (5)
  • lambda_adv = 1e-4
    Loss weight for adversarial distillation in Eq. 7.
  • lambda_ssl = 1e-4
    Loss weight for self-supervised objective in Eq. 7.
  • tau = 0.07
    Temperature in contrastive noise-model pre-training, Eq. 3.
  • lambda (noise pre-training) = 1
    Weight on contrastive loss in Eq. 4.
  • D (alignment dimension)
    Shared embedding dimension for teacher features; chosen by hand but not reported in the paper.
assumptions (4)
  • domain assumption Teacher PFMs (UNI-v2, Phikon-v2, Virchow-v2) provide useful breast-cancer representations.
    The entire distillation pipeline treats frozen teacher features as ground truth for morphology, Section 2.1.
  • ad hoc to paper Site labels in TCGA-BRCA are a valid proxy for non-biological nuisance.
    The noise model is trained to classify site; the paper assumes this captures staining/scanner bias rather than disease-relevant signal, Section 2.2.
  • domain assumption Gradient reversal effectively suppresses nuisance features in the student.
    The method relies on the standard DANN assumption that adversarial feature alignment removes predictive domain information, Section 2.2, Eq. 5.
  • domain assumption External cohorts (Yale HER2, SLN-Breast, BRACS) are representative breast-cancer evaluation sets.
    Evaluation claims are built on these three datasets being adequate for generalization, Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Harnessing Adversarial Distillation to Customise Debiased, Disease-Specific Pathology Foundation Models for Breast Cancer." pith.science (2026). https://pith.science/paper/XVLQQUVR

@misc{pith2026260801356,
  author       = {Pith},
  title        = {Pith review of: Harnessing Adversarial Distillation to Customise Debiased, Disease-Specific Pathology Foundation Models for Breast Cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVLQQUVR}},
  note         = {Machine review of arXiv:2608.01356}
}
abstract

Pathology foundation models (PFMs) provide strong tissue representations and have become central to digital pathology. However, deployment in disease-specific settings is limited by 1) the high computational cost of billion-parameter PFMs and 2) distribution mismatch and non-biological bias inherited from pan-cancer, multi-centre pre-training, including site-specific signatures and imbalanced disease prevalence. These factors can encourage shortcut learning and under-emphasise subtle morphology required for reliable modelling of a specific cancer type. We present SmartStu (a Smart Student), a framework to customise compact, breast-cancer-specific PFMs via distillation whilst mitigating confounding. SmartStu distils representations from multiple teacher PFMs into a lightweight student backbone. Crucially, we introduce adversarial distillation that leverages a dedicated noise model trained to predict nuisance, edge-dominated cues on the distillation set. Using this noise model as a counterexample, the adversarial objective encourages the student to recognise, yet suppress, features predictive of nuisance targets. We further incorporate multi-teacher ensemble distillation and an auxiliary self-supervised objective with artefact injection. We validate SmartStu on three external cohorts (Yale HER2, SLN-Breast, and BRACS) with multiple tiny backbones. SmartStu yields breast-cancer-specific PFMs that are over $30\times$ smaller than general PFMs whilst largely preserving, and sometimes improving, downstream performance measured by balanced accuracy (bAcc) and AUC. Code is available at https://github.com/zwchen03/advDistall.

Figures

Figures reproduced from arXiv: 2608.01356 by the authors.

Figure 1
Figure 1. Overview of SmartStu: multi-teacher feature distillation into a compact student, augmented with adversarial (noise) distillation to suppress site-related confounding, and an auxiliary self-supervised objective with artefact injection. project them to a normalised embedding z. We optimise a hybrid objective that combines contrastive learning [16] with cross-entropy (CE(·)) site classification: Lctr = X i∈I −1 |P(i)| … view at source ↗
Figure 2
Figure 2. Efficiency and interpretability analysis of the proposed method. (A) Para￾meter count comparison between student and teacher models. Abbreviations: MN (MobileNet-V3), R (ResNet), and TV (TinyViT). (B) Performance-efficiency trade-off (Balanced Accuracy vs. GFLOPs), where bubble size is proportional to the parameter count. (C) Slide-level attention heatmaps comparison on a sample from the Yale HER2 test set. (D) Fine… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 12 canonical work pages

  1. [1]

    NPJ Precision Oncology8(1), 151 (2024)

    Asadi-Aghbolaghi, M., Darbandsari, A., Zhang, A., Contreras-Sanz, A., Boschman, J., Ahmadvand, P., Köbel, M., Farnell, D., Huntsman, D.G., Churg, A., et al.: Learning generalizable ai models for multi-center histopathology image classifica- tion. NPJ Precision Oncology8(1), 151 (2024)

    work page 2024

  2. [2]

    Database2022, baac093 (2022)

    Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., et al.: Bracs: A dataset for breast carcinoma subtyping in h&e histology images. Database2022, baac093 (2022)

    work page 2022

  3. [3]

    The Cancer Imaging Archive (2019), dOI: 10.7937/TCIA.2019.3XBN2JCC

    Campanella, G., Hanna, M.G., Brogi, E., Fuchs, T.J.: Breast meta- stases to axillary lymph nodes. The Cancer Imaging Archive (2019), dOI: 10.7937/TCIA.2019.3XBN2JCC

    work page doi:10.7937/tcia.2019.3xbn2jcc 2019

  4. [4]

    Nature medicine25(8), 1301–1309 (2019)

    Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical- grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine25(8), 1301–1309 (2019)

    2019

  5. [5]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  6. [6]

    Nature Medicine (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Chen, B., Zhang, A., Shao, D., Song, A.H., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024)

    work page 2024

  7. [7]

    Nature medicine24(10), 1559–1567 (2018)

    Coudray,N.,Ocampo,P.S.,Sakellaropoulos,T.,Narula,N.,Snuderl,M.,Fenyö,D., Moreira, A.L., Razavian, N., Tsirigos, A.: Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nature medicine24(10), 1559–1567 (2018)

    work page 2018

  8. [8]

    Modern Pathology35(1), 44–51 (2022) 10 Zhiwei Chen, Yang Hu and Yuxiang Xiao et al

    Farahmand, S., Fernandez, A.I., Ahmed, F.S., Rimm, D.L., Chuang, J.H., Reis- enbichler, E., Zarringhalam, K.: Deep learning trained on hematoxylin and eosin tumor region of interest predicts her2 status and trastuzumab treatment response in her2+ breast cancer. Modern Pathology35(1), 44–51 (2022) 10 Zhiwei Chen, Yang Hu and Yuxiang Xiao et al

    work page 2022

Show all 28 references
  1. [9]

    The Cancer Imaging Archive (2022), dOI: 10.7937/E65C-AM96

    Farahmand, S., Fernandez, A.I., Ahmed, F.S., Rimm, D.L., Chuang, J.H., Reis- enbichler, E., Zarringhalam, K.: HER2 and trastuzumab treatment response H&E slides with tumour ROI annotations (version 3). The Cancer Imaging Archive (2022), dOI: 10.7937/E65C-AM96

    2022 doi

  2. [10]

    In: Medical Image Computing and Computer- Assisted Intervention – MICCAI 2025

    Filiot, A., Dop, N., Tchita, O., Riou, A., Dubois, R., Peeters, T., Valter, D., Scal- bert, M., Saillard, C., Robin, G., et al.: Distilling foundation models for robust and efficient models in digital pathology. In: Medical Image Computing and Computer- Assisted Intervention –...

    2025

  3. [11]

    arXiv preprint arXiv:2409.09173 (2024)

    Filiot, A., Jacob, P., Mac Kain, A., Saillard, C.: Phikon-v2: A large, public feature extractor for biomarker prediction. arXiv preprint arXiv:2409.09173 (2024)

    2024 arXiv

  4. [12]

    Journal of machine learning research17(59), 1–35 (2016)

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of machine learning research17(59), 1–35 (2016)

    2016

  5. [13]

    Nature communications12(1), 4423 (2021)

    Howard, F.M., Dolezal, J., Kochanny, S., Schulte, J., Chen, H., Heij, L., Huo, D., Nanda, R., Olopade, O.I., Kather, J.N., et al.: The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nature communications12(1), 4423 (2021)

    2021

  6. [14]

    In: International conference on machine learning

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: International conference on machine learning. pp. 2127–2136. PMLR (2018)

    2018

  7. [15]

    ACM Computing Surveys57(11), 1–37 (2025)

    Jahanifar, M., Raza, M., Xu, K., Vuong, T.T.L., Jewsbury, R., Shephard, A., Zamanitajeddin, N., Kwak, J.T., Raza, S.E.A., Minhas, F., et al.: Domain gen- eralization in computational pathology: survey and guidelines. ACM Computing Surveys57(11), 1–37 (2025)

    2025

  8. [16]

    Advances in neural information processing systems33, 18661–18673 (2020)

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems33, 18661–18673 (2020)

    2020

  9. [17]

    arXiv preprint arXiv:2411.05489 (2024)

    Kömen, J., Marienwald, H., Dippel, J., Hense, J.: Do histopathological found- ation models eliminate batch effects? a comparative study. arXiv preprint arXiv:2411.05489 (2024)

    2024 arXiv

  10. [18]

    arXiv preprint arXiv:2501.15724 (2025)

    Li, D., Wan, G., Wu, X., Wu, X., Nirmal, A.J., Lian, C.G., Sorger, P.K., Semenov, Y.R., Zhao, C.: A survey on computational pathology foundation models: Data- sets, adaptation strategies, and evaluation tasks. arXiv preprint arXiv:2501.15724 (2025)

    2025 arXiv

  11. [19]

    In: 2020 international joint conference on neural networks (IJCNN)

    Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1–7. IEEE (2020)

    2020

  12. [20]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  13. [21]

    IEEE transactions on biomedical engineering61(5), 1400–1411 (2014)

    Veta, M., Pluim, J.P., Van Diest, P.J., Viergever, M.A.: Breast cancer histopatho- logy image analysis: A review. IEEE transactions on biomedical engineering61(5), 1400–1411 (2014)

    2014

  14. [22]

    Nature medicine30(10), 2924–2935 (2024)

    Vorontsov, E., Bozkurt, A., Casson, A., Shaikovski, G., Zelechowski, M., Sever- son, K., Zimmermann, E., Hall, J., Tenenholtz, N., Fusi, N., et al.: A founda- tion model for clinical-grade computational pathology and rare cancers detection. Nature medicine30(10), 2924–2935 (2024)

    2024

  15. [23]

    arXiv preprint arXiv:2504.04045 (2025)

    Xiong, C., Chen, H., Sung, J.J.: A survey of pathology foundation model: Progress and future directions. arXiv preprint arXiv:2504.04045 (2025)

    2025 arXiv

  16. [24]

    Nature630(8015), 181–188 (2024) Adversarial Distillation for Debiased Breast Cancer Foundation Models 11

    Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Naumann, T., Wong, C., Gero, Z., González, J., Gu, Y., et al.: A whole-slide foundation model for digital pathology from real-world data. Nature630(8015), 181–188 (2024) Adversarial Distillation for Debiased Breast Cancer Fou...

    2024

  17. [25]

    Nature Communications (2025)

    Xu, Y., Wang, Y., Zhou, F., Ma, J., Jin, C., Yang, S., Li, J., Zhang, Z., Zhao, C., Zhou, H., et al.: A multimodal knowledge-enhanced whole-slide pathology found- ation model. Nature Communications (2025)

    2025

  18. [26]

    arXiv preprint arXiv:2502.06750 (2025)

    Zhang, A., Jaume, G., Vaidya, A., Ding, T., Mahmood, F.: Accelerating data processing and benchmarking of ai models for pathology. arXiv preprint arXiv:2502.06750 (2025)

    2025 arXiv

  19. [27]

    arXiv preprint arXiv:2111.07832 (2021)

    Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)

    2021 arXiv

  20. [28]

    arXiv preprint arXiv:2408.00738 (2024)

    Zimmermann, E., Vorontsov, E., Viret, J., Casson, A., Zelechowski, M., Shaikovski, G., Tenenholtz, N., Hall, J., Fuchs, T., Fusi, N., Liu, S., Severson, K.: Virchow2: Scaling self-supervised mixed-magnification models in pathology. arXiv preprint arXiv:2408.00738 (2024)

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.