Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss

T0 review · 3 major / 6 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 Pathology foundation models are scanner sensitive: the same tissue digitized by different commercial scanners yields different model predictions, and a contrastive projection loss called ScanGen, applied during task fine-tuning, cuts this…

desk verdict Useful benchmark, but the headline CoV improvements may be a metric artifact—worth refereeing with mandatory logit-statistics and confidence-interval checks. read the letter →

arxiv 2507.22092 v1 pith:W4BLC4LP submitted 2025-07-29 q-bio.QM cs.AIcs.CVeess.IVq-bio.TO

Gianluca Carloni , Biagio Brattoli , Seongho Keum , Jongchan Park , Taebum Lee , Chang Ho Ahn , Sergio Pereira This is my paper
classification q-bio.QMcs.AIcs.CVeess.IVq-bio.TO
keywords computationalpathologyscannerbiasfoundationmodelswholeslideimagesmultipleinstancelearningcontrastiveEGFRmutationpredictiondomaingeneralization
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 argues that pathology foundation models, despite being promoted for domain generalization, are scanner sensitive: the same tissue digitized by different commercial scanners yields noticeably different model predictions. The authors benchmark five published foundation models on 323 lung-cancer specimens, each scanned by six scanners at 40x plus one scanner at 20x, and measure scanner bias as the coefficient of variation (CoV) across same-specimen predictions, finding CoV values from 0.324 to 0.450. They then propose ScanGen, a contrastive loss applied to a small projection head during task fine-tuning, which pulls same-specimen different-scanner embeddings together and pushes same-scanner different-specimen embeddings apart. On the held-out test set, ScanGen cuts scanner CoV by 8% to 69% depending on the foundation model, while EGFR mutation AUC is retained or slightly improved for every model tested. If this holds, scanner robustness can be added on top of frozen embeddings at fine-tuning time, without retraining the foundation model itself.

What carries the argument

The central object is the ScanGen contrastive loss, a reworking of the classic contrastive loss from the paper's reference [7], defined on a learned projection network $h(\cdot)$ inserted before the MIL aggregator. It combines an attraction term that minimizes cosine distance between embeddings of the same specimen scanned by different scanners and a repulsion term that pushes apart embeddings of different specimens scanned by the same scanner, weighted by the hyper-parameters $\alpha$, $r$, and $\lambda$, and summed with the task's cross-entropy loss. Its job is to reorganize the frozen foundation-model embedding space around specimen identity rather than scanner identity, so downstream classification becomes insensitive to acquisition hardware. The companion metric, scanner CoV, the coefficient of variation of logit predictions across scanners for the same specimen, is what makes the bias measurable.

What would settle it

Train the ScanGen branch on paired slides from a subset of the six scanners and evaluate CoV on a scanner that never appeared in training: if the CoV reduction does not transfer to the unseen scanner, the claimed scanner invariance is only interpolation among familiar devices. A second check: on a cohort where each specimen is scanned by exactly one scanner, ScanGen has no real pairs to learn from, so the method would have to rely on synthetic pairs; if that variant fails, the paired-data requirement is the binding constraint.

Watch

Extended reading notes

Core claim

On a multi-scanner benchmark of 323 lung-cancer specimens (2,261 whole slide images), five pathology foundation models embed the same tissue differently depending on the scanner that digitized it: the scanner-level coefficient of variation in EGFR-mutation logits ranges from 0.324 for H-Optimus-0 to 0.450 for Gigapath. The paper's ScanGen loss reduces this bias on a held-out set of 224 specimens while improving the downstream EGFR AUC for every embedding tested; for example, Phikon's CoV falls from 0.389 to 0.121 with AUC rising from 0.738 to 0.768, and the gains also reduce disagreement across magnifications. The authors further show that the effect grows monotonically with the number of scanner types seen in training, with three scanners already yielding a 27% CoV improvement against six test scanners, and that ScanGen transfers to several MIL aggregators including GAP, AB-MIL, DS-MIL, and SlotMIL.

Load-bearing premise

The method needs paired whole-slide images of the same specimen captured by different scanners for its training branch, yet only 54 such specimens are used, and in routine clinical practice a slide is usually scanned once, so this paired data is itself an expensive prerequisite whose sufficiency is assumed.

Editorial extensions

If this is right

  • ScanGen can be added to any frozen-foundation-model MIL pipeline as a projection head, so existing clinical models can be made more scanner-robust without retraining the large foundation model.
  • Scanner robustness is not bought at the cost of task performance: on the paper's numbers, every tested foundation model either keeps or slightly improves EGFR AUC while lowering scanner CoV.
  • The benefit scales with scanner diversity in training: more scanner types mean lower CoV, with diminishing returns around five scanners.
  • The loss also reduces magnification disagreement, covering a second acquisition-variation axis beyond hardware vendor.
  • The effect holds across multiple MIL aggregators, so the mitigation is not tied to one aggregation choice.

Reading between the lines

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

  • The strongest untested claim implied here is generalization to scanners absent from training; the paper's ablation varies which subsets of the same six scanners are used, so evaluating on a seventh scanner never seen in training would be a stricter test of true scanner invariance.
  • Because the attraction term effectively injects specimen identity as a training signal, part of the measured AUC gain could come from specimen-pair structure rather than pure biological signal; the specimen-level held-out split partially addresses this, but an analysis ablating the attract term alone would settle it.
  • Synthetic pairing, simulating scanner appearance shifts through stain or optical augmentation, could remove the need to physically re-scan specimens and would make the method usable where only single scans per specimen exist.
  • The approach is not specific to EGFR: the same projection-head loss should transfer to other WSI-level tasks built on frozen foundation models, such as survival or immune-cell scoring, where scanner bias would similarly corrupt hospital-to-hospital deployment.
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 / 6 minor

Summary. The paper addresses scanner sensitivity of pathology foundation models (FMs). It introduces a benchmark using a multi-scanner dataset of 323 specimens each scanned with six scanners at 40× and one at 20×, and proposes ScanGen, a contrastive loss applied to a projection head during MIL fine-tuning for EGFR mutation prediction. The authors report that five state-of-the-art FMs exhibit substantial scanner CoV on held-out specimens, and that ScanGen reduces this CoV while maintaining or improving EGFR AUC. An ablation shows that performance improves with more scanner types used in training, and the method extends to several MIL aggregators.

Significance. The problem is timely and important: scanner bias in pathology FMs is a real barrier to clinical deployment, and the paper proposes a practical mitigation that works at the feature-extraction level. The experimental design has notable strengths: specimen-level splits, a held-out test set of 224 specimens with paired multi-scanner slides, a fair comparison (both baseline and ScanGen include the same projection head), and an ablation varying the number and combination of scanners. The observation that ScanGen transfers to unseen specimens and new scanner combinations is valuable evidence. If the mitigation claim is robust, the method could be widely applicable. However, the central quantitative metric (CoV of logits) is problematic and the results lack uncertainty quantification, so the significance is conditional on addressing these issues.

major comments (3)
  1. [Section 2.1, Eq. (1), Table 1] The CoV metric on logits is not a valid measure of scanner sensitivity for unbounded, signed values. CoV = std/|mean| conflates dispersion with the magnitude of the mean logit. A model that produces more extreme logits (e.g., shifting from logits near 0 to logits near ±2) will show a lower CoV even if the absolute scanner-induced standard deviation is unchanged or larger. Since ScanGen applies an attraction term that can change the scale of representations, the reported CoV reductions (e.g., Phikon from 0.389 to 0.121) could be an artifact of a mean shift rather than a genuine reduction in scanner-induced variability. The paper does not report the mean logit or the raw standard deviation for any model. AUC is rank-based and scale-invariant, so the AUC improvements do not control for this confound. Please provide the raw standard deviation across scanners and the mean logit values for each model, or use a metric that is invariant to monotonic transformations of the logits (e.g., rank-based agreement or variance after normalizing the mean). Without this, the central mitigation claim is not fully supported.
  2. [Table 1, Section 3.2] The results are reported for a single run with no confidence intervals, standard errors, or multiple seeds. Given that the ScanGen training set contains only 54 specimens, the variability of both CoV and AUC estimates could be substantial. The paper should report bootstrap confidence intervals or results over multiple random seeds, and ideally a paired significance test for the CoV reductions. This is necessary to establish that the observed improvements are not due to chance.
  3. [Table 1, Gigapath row] The magnification CoV for Gigapath increases under ScanGen from 0.174 to 0.193 (+10.9%). This contradicts the general claim in Section 3.2 that "the ScanGen loss also helps mitigate a disagreement in predictions when different magnifications are used." Please either provide an explanation for this exception or qualify the claim to acknowledge that magnification robustness is not consistently improved.
minor comments (6)
  1. [Table 1 caption] The caption reads "Coefficient of EGFR Variation" but should be "Coefficient of Variation"; EGFR is the task, not the metric.
  2. [Eq. (1) and Section 2.1] The notation \hat{y}^{s_1}_m is used for logit predictions, but earlier \hat{y} is defined as the final prediction. Clarify that CoV is computed on the logits (pre-sigmoid) and define the notation precisely.
  3. [Section 2.2, Eq. (2)] The ScanGen loss is defined for a pair of specimens and two scanners, but the averaging over all valid pairs in a batch is only described in Section 2.3. Please make the batch-level averaging explicit in the loss definition or reference Section 2.3.
  4. [Section 3.1] The sentences "The specimens were scanned at 40× magnification. The specimens were also scanned with Leica Aperio GT450 at 20× magnification." are repetitive; consider rewriting to state that each specimen was scanned with the six scanners at 40× and additionally with the GT450 at 20×.
  5. [Section 3.3] The claim about which scanner combinations are best/worst would be easier to verify with a supplementary table listing the specific combinations and their CoV values; the current text only provides a summary.
  6. [References] Reference [20] (H-Optimus-0) is a GitHub link; if a peer-reviewed or arXiv version exists, it should be cited for permanence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ScanGen improvements are measured on held-out specimens and scanner combinations, and the loss is not definitionally identical to the CoV metric.

full rationale

The load-bearing empirical claims are that pathology FMs are scanner sensitive (Table 1 baselines) and that ScanGen reduces scanner CoV on a held-out test set while retaining or improving EGFR AUC. The ScanGen loss (Eq. 2) operates on patch embeddings via cosine distance, whereas the CoV metric (Eq. 1) is computed on final logit predictions after MIL aggregation and classification; the two are therefore not the same quantity by construction. The paper explicitly reserves 54 specimens for ScanGen training, 45 for validation, and 224 for the held-out test set, so the reported CoV reductions (e.g., Phikon 0.389 to 0.121) are not training-loss artifacts. Section 3.3 further evaluates generalization to scanner combinations not used in training, which is an independent check. The self-references in the paper (e.g., SlotMIL from author Keum, and a prior Lunit study on tumor-infiltrating lymphocytes) are contextual or peripheral and carry no load-bearing justificatory weight. The skeptic concern that CoV of logits may be confounded by logit scale is a question of metric validity, not of circular derivation: it does not show that the claimed reduction reduces to the paper's inputs by construction. Accordingly, no circular step meeting the quoted-evidence standard is present.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. Free parameters are the standard loss weights and projection network width tuned on validation data. The key assumptions are the sufficiency of the small paired training set and the meaningfulness of CoV as a bias metric.

free parameters (4)
  • alpha (α) = 0.12 to 0.21
    Weight balancing attract and repel terms in ScanGen loss; tuned on validation set.
  • radius (r) = 0.90 to 1.10
    Margin for the repulsion term; tuned on validation.
  • lambda (λ) = 0.5 to 1.0
    Balance between CE and ScanGen loss; tuned on validation.
  • MLP hidden dimension = 48 to 96
    Projection network width, chosen per foundation model feature size.
assumptions (3)
  • domain assumption The multi-scanner dataset is representative of real-world scanner variation and the 54-specimen ScanGen training subset is sufficient to learn scanner invariance that transfers to new specimens.
    The central claim that ScanGen mitigates scanner bias depends on this subset generalizing; no analysis of training size sensitivity is given beyond scanner count.
  • domain assumption CoV of raw logits is a meaningful measure of scanner bias even when mean logits approach zero.
    CoV divides by |mean| and can explode near decision boundaries; the paper does not report how many specimens have near-zero mean logits.
  • domain assumption The paired same-specimen multi-scanner scans exist for all training specimens, which is rare in routine clinical practice.
    ScanGen requires multiple scans of the same specimen; if such pairs cannot be obtained, the method cannot be applied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss." pith.science (2026). https://pith.science/paper/W4BLC4LP

@misc{pith2026250722092,
  author       = {Pith},
  title        = {Pith review of: Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4BLC4LP}},
  note         = {Machine review of arXiv:2507.22092}
}
read the original abstract

Computational pathology (CPath) has shown great potential in mining actionable insights from Whole Slide Images (WSIs). Deep Learning (DL) has been at the center of modern CPath, and while it delivers unprecedented performance, it is also known that DL may be affected by irrelevant details, such as those introduced during scanning by different commercially available scanners. This may lead to scanner bias, where the model outputs for the same tissue acquired by different scanners may vary. In turn, it hinders the trust of clinicians in CPath-based tools and their deployment in real-world clinical practices. Recent pathology Foundation Models (FMs) promise to provide better domain generalization capabilities. In this paper, we benchmark FMs using a multi-scanner dataset and show that FMs still suffer from scanner bias. Following this observation, we propose ScanGen, a contrastive loss function applied during task-specific fine-tuning that mitigates scanner bias, thereby enhancing the models' robustness to scanner variations. Our approach is applied to the Multiple Instance Learning task of Epidermal Growth Factor Receptor (EGFR) mutation prediction from H\&E-stained WSIs in lung cancer. We observe that ScanGen notably enhances the ability to generalize across scanners, while retaining or improving the performance of EGFR mutation prediction.

Figures

Figures reproduced from arXiv: 2507.22092 by the authors.

Figure 1
Figure 1. Left: Existing pathology Foundation Models suffer from scanner bias. We show a 2D projection of WSIs (dots) from 4 scanners (colors) using UMAP based on H￾Optimus-0 embeddings before and after ScanGen. "SLIDE−→" indicates the slides of the same specimen from different scanners. Right: ScanGen improves both generalization and performance. new capabilities to clinical practice, such as prediction of genetic alteration… view at source ↗
Figure 2
Figure 2. Our method. Left: Data are projected to a new embedding space where scanner bias is removed via our contrastive ScanGen loss. ScanGen is trained on paired WSIs of the same specimen acquired with different scanners and can be easily integrated with any existing method by acting before the MIL aggregator. Right: ScanGen attracts the representations of same-specimen different-scanner pairs, while repelling same-scanner… view at source ↗
Figure 3
Figure 3. Our multi-scanner dataset projected to 2D using various FMs. Each dot is a WSI embedding from the test set by averaging FM embeddings across patches and projecting them using UMAP. The color indicates the scanner. We selected 4 out of 7 scanners to reduce clutter. A single slide is indicated with "SLIDE−→" to emphasize the location of the slides of the same specimen from different scanners in each plot. highlights t… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Left: Scanner agreement improves with the number of scanners used for training ScanGen. Vertical bars indicate the standard deviation across scanner combinations. Right: Examples of the same specimen digitized using different scanners. 3.4 Extension to Other MIL Method…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Enabling clinical use of foundation models for computational pathology

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Novel robustness losses added during downstream training on foundation-model features from pathology slides improve both robustness to technical variation and classification accuracy.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Albertina, B., Watson, M., Holback, C., Jarosz, R., Kirk, S., Lee, Y., Rieger- Christ, K., Lemmerman, J.: The cancer genome atlas lung adenocarcinoma collec- tion (tcga-luad) (2016)

    work page 2016

  2. [2]

    arXiv preprint arXiv:2206.10573 (2022)

    Campanella, G., Ho, D., Häggström, I., Becker, A.S., Chang, J., Vanderbilt, C., Fuchs, T.J.: H&e-based computational biomarker enables universal egfr screening for lung adenocarcinoma. arXiv preprint arXiv:2206.10573 (2022)

    arXiv 2022

  3. [3]

    Nature Medicine30(3), 850–862 (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Song, A.H., Chen, B., Zhang, A., Shao, D., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine30(3), 850–862 (2024)

    2024

  4. [4]

    https://doi.org/10.7937/K9/TCIA.2018.PAT12TBS

    (CPTAC), N.C.I.C.P.T.A.C.: The clinical proteomic tumor analysis consor- tium lung adenocarcinoma collection (cptac-luad) (version 12) [data set] (2018). https://doi.org/10.7937/K9/TCIA.2018.PAT12TBS

    work page doi:10.7937/k9/tcia.2018.pat12tbs 2018

  5. [5]

    Journal of Pathology Informatics14 (2023)

    Duenweg, S.R., Bobholz, S.A., Lowman, A.K., Stebbins, M.A., Winiarz, A., Nath, B., Kyereme, F., Iczkowski, K.A., LaViolette, P.S.: Whole slide imaging (wsi) scan- ner differences influence optical and computed properties of digitized prostate can- cer histology. Journal of Pathology Informatics14 (2023)

    work page 2023

  6. [6]

    medRxiv (2023)

    Filiot, A., Ghermi, R., Olivier, A., Jacob, P., Fidon, L., Camara, A., Mac Kain, A., Saillard, C., Schiratti, J.B.: Scaling self-supervised learning for histopathology with masked image modeling. medRxiv (2023)

    work page 2023

  7. [7]

    In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06)

    Hadsell, R., Chopra, S., LeCun, Y.: Dimensionality reduction by learning an in- variant mapping. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06). vol. 2, pp. 1735–1742. IEEE (2006)

    work page 2006

  8. [8]

    In: International conference on machine learning

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

    work page 2018

Show all 26 references
  1. [9]

    arXiv preprint arXiv:2501.18055 (2025)

    de Jong, E.D., Marcus, E., Teuwen, J.: Current pathology foundation models are unrobust to medical center differences. arXiv preprint arXiv:2501.18055 (2025)

    2025 arXiv

  2. [10]

    Frontiers in Medicine8, 746307 (2021)

    Kang, H., Luo, D., Feng, W., Zeng, S., Quan, T., Hu, J., Liu, X.: Stainnet: a fast and robust stain normalization network. Frontiers in Medicine8, 746307 (2021)

    2021

  3. [11]

    Nature cancer1(8), 789–799 (2020)

    Kather, J.N., Heij, L.R., Grabsch, H.I., Loeffler, C., Echle, A., Muti, H.S., Krause, J., Niehues, J.M., Sommer, K.A., Bankhead, P., et al.: Pan-cancer image-based detection of clinically actionable genetic alterations. Nature cancer1(8), 789–799 (2020)

    2020

  4. [12]

    arXiv preprint arXiv:2311.17466 (2023)

    Keum, S., Kim, S., Lee, S., Lee, J.: Slot-mixup with subsampling: A simple regu- larization for wsi classification. arXiv preprint arXiv:2311.17466 (2023)

    2023 arXiv

  5. [13]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 14318–14328 (2021)

    2021

  6. [14]

    In: 2009 IEEE international symposium on biomedical imaging: from nano to macro

    Macenko, M., Niethammer, M., Marron, J.S., Borland, D., Woosley, J.T., Guan, X., Schmitt, C., Thomas, N.E.: A method for normalizing histology slides for quan- titative analysis. In: 2009 IEEE international symposium on biomedical imaging: from nano to macro. pp. 1107–1110. IE...

    2009

  7. [15]

    arXiv preprint arXiv:1802.03426 (2018)

    McInnes, L., Healy, J., Melville, J.: Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  8. [16]

    Journal of Clinical Oncology40(17), 1916– 1928 (2022) Pathology FMs are Scanner Sensitive: Benchmark & Mitigation 11

    Park, S., Ock, C.Y., Kim, H., Pereira, S., Park, S., Ma, M., Choi, S., Kim, S., Shin, S., Aum, B.J., et al.: Artificial intelligence–powered spatial analysis of tumor- infiltrating lymphocytes as complementary biomarker for immune checkpoint in- hibition in non–small-cell lung...

    2022

  9. [17]

    medRxiv (2023)

    Pinto, D.G., Bychkov, A., Tsuyama, N., Fukuoka, J., Eloy, C.: Exploring the adop- tion of digital pathology in clinical settings-insights from a cross-continent study. medRxiv (2023)

    2023

  10. [18]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Ramanathan, V., Pati, P., McNeil, M., Martel, A.L.: Ensemble of prior-guided expert graph models for survival prediction in digital pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 262–272 (2024)

    2024

  11. [19]

    IEEE Computer graphics and applications21(5), 34–41 (2001)

    Reinhard, E., Adhikhmin, M., Gooch, B., Shirley, P.: Color transfer between im- ages. IEEE Computer graphics and applications21(5), 34–41 (2001)

    2001

  12. [20]

    Saillard, C., Jenatton, R., Llinares-López, F., Mariet, Z., Cahané, D., Durand, E., Vert, J.P.: H-optimus-0 (2024), https://github.com/bioptimus/releases/tree/main/models/h-optimus/v0

    2024

  13. [21]

    In: 2019 Ieee 16th international symposium on biomedical imaging (Isbi 2019)

    Shaban, M.T., Baur, C., Navab, N., Albarqouni, S.: Staingan: Stain style trans- fer for digital histological images. In: 2019 Ieee 16th international symposium on biomedical imaging (Isbi 2019). pp. 953–956. IEEE (2019)

    2019

  14. [22]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Teichmann, M., Aichert, A., Bohnenberger, H., Ströbel, P., Heimann, T.: End-to- endlearningforimage-baseddetectionofmolecularalterationsindigitalpathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 88–98 (2022)

    2022

  15. [23]

    Nature pp

    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. Nature pp. 1–8 (2024)

    2024

  16. [24]

    arXiv preprint arXiv:2408.00380 (2024)

    Yun, J., Hu, Y., Kim, J., Jang, J., Lee, S.: Exaonepath 1.0 patch-level foundation model for pathology. arXiv preprint arXiv:2408.00380 (2024)

    2024 arXiv

  17. [25]

    BMC pulmonary medicine23(1) (2023)

    Zhao, D., Zhao, Y., He, S., Liu, Z., Li, K., Zhang, L., Zhang, X., Wang, S., Che, N., Jin, M.: High accuracy epidermal growth factor receptor mutation prediction via histopathological deep learning. BMC pulmonary medicine23(1) (2023)

    2023

  18. [26]

    arXiv preprint arXiv:2408.00738 (2024)

    Zimmermann, E., Vorontsov, E., Viret, J., Casson, A., Zelechowski, M., Shaikovski, G., Tenenholtz, N., Hall, J., Klimstra, D., Yousfi, R., et al.: Virchow2: Scal- ing 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.