Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

PhenoBench: A Comprehensive Benchmark for Cell Phenotyping

T0 review · 4 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 Benchmark shows pathology AI drops to 0.20 F1 on fine-grained cell typing

desk verdict Useful new benchmark, but the headline claim about task difficulty needs a human-upper-bound control before it can be taken at face value. read the letter →

arxiv 2507.03532 v6 pith:BG4ONOPX submitted 2025-07-04 cs.CV

Claudia Winklmayr , Jerome Luescher , Nora Koreuber , Jannik Franzen , Fabian H. Reith , Elias Baumann , Christian M. Schuerch , Dagmar Kainmueller
show 1 more author
Josef Lorenz Rumberger
This is my paper
classification cs.CV
keywords cellphenotypingfoundationmodelshistopathologyH&Estainingbenchmarkdomainshiftmultiplexedimagingcolorectalcancer
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 introduces PhenoBench, a benchmark for cell phenotyping in H&E stained tissue, built around a new dataset called PhenoCell that labels 88 million cells into 14 cell types. The authors benchmark seven pathology foundation models and a supervised baseline, and find that models scoring above 0.70 macro F1 on established datasets such as Lizard and PanNuke fall to scores as low as 0.20 on PhenoCell. They argue this gap shows that previous benchmarks are nearing saturation and that PhenoCell captures a harder task, including fine-grained classes and medically meaningful domain shifts by tumor type and cancer stage. The paper also reports that foundation models are more data-efficient and more robust to technical domain shifts than the supervised baseline, but more sensitive to biological domain shifts.

What carries the argument

The load-bearing mechanism is the pairing of the PhenoCell dataset with a standardized benchmarking pipeline. PhenoCell supplies dense panoptic masks for 88 million cells in H&E images, with cell-type labels derived from CODEX antibody marker expression, clustered by X-shift and refined through manual expert review. The pipeline freezes each foundation model's encoder and trains two decoding heads, a linear probe and a UNetR decoder, for dense prediction, then measures macro F1 across four datasets and multiple splits. The score collapse on PhenoCell relative to saturated benchmarks is what carries the paper's argument that the field needs a harder evaluation standard.

What would settle it

Have expert pathologists re-annotate a random sample of PhenoCell cells directly from H&E slides and re-run the benchmark on the cleaned labels; if scores rise substantially above 0.28, the benchmark's difficulty is partly an artifact of its noisy automatic labels. Alternatively, if a model trained on high-quality manual H&E annotations reaches high F1 on PhenoCell, the claim that the task is uniquely hard weakens.

Watch

Extended reading notes

Core claim

The central claim is that PhenoCell, a dataset derived from multiplexed CODEX imaging of colorectal cancer tissue, constitutes a fundamentally harder cell phenotyping benchmark than existing ones. Its 14 granular cell types are highly imbalanced, with tumor cells outnumbering natural killer cells 153-to-1, and its ground-truth labels include morphologically ambiguous classes such as dendritic and NK cells. Evaluated under three splits—random, tumor-type, and tumor-stage—all seven foundation models score between 0.20 and 0.28 macro F1, while the same models exceed 0.70 on Lizard and PanNuke. The authors conclude that PhenoCell establishes a more challenging task not captured by previous benchmarks and should become a prime evaluation asset for both foundation models and supervised methods.

Load-bearing premise

The ground-truth labels from automatic CODEX segmentation and clustering are accurate enough that low foundation-model scores on PhenoCell reflect task difficulty rather than label noise or biologically ambiguous cell types.

Editorial extensions

If this is right

  • Future cell phenotyping evaluations should include PhenoCell, since established benchmarks like Lizard and PanNuke are already near saturation for foundation models.
  • Foundation models beat the supervised HoVer-NeXt baseline in low-data regimes, by up to 0.14 macro F1 when trained on 5% of Lizard data, so they are a practical choice for annotation-scarce projects.
  • On medical domain shifts—tumor type and cancer stage—foundation models lose relatively more performance than the supervised baseline, so biological generalization needs dedicated attention.
  • The UNetR decoder consistently outperforms linear probing for dense cell phenotype prediction, indicating that decoder design is a major factor in extraction quality.

Reading between the lines

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

  • The score gap on PhenoCell may overstate the true difficulty gap, because its automatically generated masks and morphologically ambiguous classes add label noise; a cleaned or noise-annotated version would separate task difficulty from annotation difficulty.
  • The tumor-type and tumor-stage splits are a template for clinically relevant generalization testing that could be extended to other cancer types and organs.
  • A natural next test is whether cell-type-level F1 on PhenoCell predicts performance in downstream clinical tasks such as prognosis or treatment-response prediction, which would establish the benchmark's practical value.
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 introduces PhenoBench, a cell phenotyping benchmark for H&E-stained histopathology, centered on a new dataset called PhenoCell containing 14 cell types annotated via multiplexed CODEX imaging and X-shift clustering. The authors benchmark seven pathology foundation models and the supervised HoVer-NeXt baseline under four dataset splits (random, tumor-type, tumor-stage, and a technical Lizard split), using both linear-probe and UNetR decoders. Results show strong performance on Arctique, PanNuke, and Lizard (macro F1 0.71–0.93) but much lower scores on PhenoCell (macro F1 0.28 average, with some models near 0.20). The paper argues that PhenoCell poses a substantially harder task than existing benchmarks and should be adopted as a new evaluation standard.

Significance. If the main claim holds, PhenoCell fills an important gap: existing cell phenotyping benchmarks are approaching saturation, and a demanding, publicly available benchmark with multiple medically meaningful domain splits would be valuable for progress in computational pathology. The paper's strengths include the release of code and data, evaluation of seven diverse foundation models under frozen-encoder settings, and the inclusion of biological domain shifts (tumor type and stage) that go beyond scanner/staining variation. The authors also disclose the known limitation of automatic segmentation quality. However, the central claim that low F1 on PhenoCell indicates task difficulty rather than label ambiguity is not yet supported, because the ground truth was derived from molecular markers that may not be visible in H&E and no human-achievable upper bound is established. The benchmark's utility for future model comparison depends on resolving this label-reliability question, so the result is significant but currently conditional.

major comments (4)
  1. [Section 2.1] The claim that PhenoCell contains 88 million individual cells in only 109 fields of view (FoVs) is numerically implausible and, if taken at face value, would imply approximately 800,000 cells per FoV. This is three orders of magnitude larger than the per-FoV counts reported for Lizard (about 1,700 cells per FoV) and appears inconsistent with typical whole-slide image resolutions. The authors should report the per-FoV cell count distribution, the tile size used for H&E staining, and clarify how the total was computed. If the number is a typo or refers to an intermediate step, it must be corrected because the paper uses '88 million' as evidence that PhenoCell is 'much larger than all previous datasets.'
  2. [Section 4 (Results) and Section 5 (Discussion)] The central claim that low PhenoCell scores indicate a 'much more challenging task not captured by previous benchmarks' conflates task difficulty with label modality mismatch. PhenoCell's ground-truth labels are derived from CODEX protein expression, and classes such as NK cells, dendritic cells, B cells, T cells, and non-nuclear structures (nerves, vasculature, adipocytes) may be impossible to identify from H&E morphology alone. The paper acknowledges in Section 5 that NK and dendritic cells are morphologically ambiguous, but it never quantifies the extent to which a perfect H&E-based model could approach the reported scores. Without a human expert upper bound (e.g., a panel of pathologists annotating a subset of PhenoCell tiles from H&E) or an inter-observer agreement study, the comparison to Lizard and PanNuke (whose labels were created manually from H&E) is not apples-to-apples. The authors should provide such a calibration or explicitly rephrase the claim as 'FMs perform poorly on a label space defined by molecular markers,' which is weaker and less actionable.
  3. [Section 5] The paper states that 'the segmentation masks of PhenoCell were created via an automatic segmentation tool, thus their quality is lower than that of other datasets,' but it does not quantify this label noise. Since the benchmark's utility depends on the reliability of ground truth, the authors should report (i) the fraction of cells or FoVs that were flagged as false merges or false negatives during the visual quality-control step, (ii) a sample of manual re-annotation agreement on a subpopulation, and (iii) per-class reliability statistics, especially for NK and dendritic cells. Without this, the 0.20–0.28 F1 range could be substantially inflated or deflated by label noise rather than by model capability. At minimum, the paper should report scores with and without the ambiguous classes to show the robustness of the headline result.
  4. [Section 4 (Experimental Setup)] All experiments appear to be run with a single random seed and no repeated trials, yet the paper draws conclusions from small differences, e.g., a 0.03 F1 drop in the Tumor-Type-Split or 0.02–0.05 gaps between models. Given the acknowledged label noise, these differences may be within run-to-run variance. The authors should provide standard deviations across at least three seeds for the main benchmark tables, or otherwise state that such differences are not statistically significant. This is particularly important for the comparisons in Figure 3 and Figure 4b, where the 'advantage' of foundation models is often on the order of 0.03–0.06 F1.
minor comments (5)
  1. [Figure 4b] The caption contains a typo: 'ar 50%' should read 'at 50%'.
  2. [References] References [4] (TITANv) and [9] (PanNuke) list the same arXiv identifier 2003.10778. The TITANv reference should be updated to its correct identifier (2404.xxxx or similar), as the current citation is misleading.
  3. [Figure 3 and Section 4 text] The text says 'In PhenoCell, we evaluate robustness to domain shifts using the technical (Fig. 3a) and medical data splits (Fig. 3b),' but Figure 3a is actually the Lizard Center-Split, which is the technical split, while Figure 3b and 3c are the PhenoCell medical splits. This mismatch between caption, labels, and prose should be corrected to avoid confusion.
  4. [Section 2.1] The dataset description states that 31 FoVs were excluded due to false merges or false negatives but does not specify the total number before exclusion or whether the 109 FoVs are from both TMA slides. Reporting the FoV and patient distribution across the three splits would help readers assess the domain-shift setup.
  5. [Abstract and Section 4] The abstract reports 'scores as low as 0.20' while the results section reports an average of 0.28 F1 on PhenoCell. Clarify whether 0.20 is the minimum across models/decoders, and give the exact per-model range in the text, not only in the figure, to avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PhenoCell labels come from external CODEX/X-shift data, and foundation model checkpoints are frozen public weights, so the benchmark results are not equivalent to their inputs.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined by a fitted parameter or by a prior claim from the same authors. The central comparison — FMs scoring 0.71–0.79 on Lizard and PanNuke but 0.20–0.28 on PhenoCell — is computed from (a) public, frozen foundation-model checkpoints trained on unrelated histopathology data, and (b) PhenoCell ground-truth masks generated externally from multiplexed CODEX imaging with X-shift clustering and expert refinement. The paper explicitly states that models were never trained on the evaluation datasets, and it discloses that HoVer-NeXt and Arctique come from the authors' own prior work; these self-citations are used only as a supervised baseline and as an additional synthetic evaluation set, so they do not bear the burden of the central claim. The admitted limitation that PhenoCell masks come from an automatic segmentation tool and that classes such as NK and dendritic cells are morphologically ambiguous is a plausible correctness risk for the claim that low F1 reflects task difficulty rather than label ambiguity, but that is a validity concern about label quality, not a circularity in which an output is equivalent to an input by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior work, and no known result is repackaged under new coordinates. The benchmark conclusions therefore stand independently of the paper's own assumptions.

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

No fitted constants or invented entities. The empirical claims rest on the fidelity of CODEX-derived labels as H&E ground truth, on the morphological distinguishability of the 14 classes, on the absence of benchmark overlap with FM pretraining data, and on the patient-disjointness of the domain splits.

assumptions (4)
  • domain assumption CODEX/X-shift cluster labels are a faithful proxy for true cell phenotypes and can be used as ground truth on the corresponding H&E images.
    Labels originate from an automatic segmentation tool plus X-shift clustering of marker expression, refined by manual review. The paper admits the segmentation masks have lower quality than manual ones, so label fidelity is assumed, not proven.
  • domain assumption H&E morphology alone is sufficient to distinguish the 14 merged cell phenotypes; otherwise low F1 reflects label ambiguity, not model failure.
    This is the interpretative premise behind the claim that PhenoCell is 'much more challenging' than prior benchmarks. It is plausible for some classes but doubtful for NK cells and dendritic cells, which the authors themselves call morphologically ambiguous.
  • domain assumption None of the benchmarked foundation models was pretrained on PhenoCell, Lizard, PanNuke, or Arctique.
    Section 2 states this was ensured, but the proprietary pretraining corpora of several models cannot be independently audited from the paper.
  • domain assumption The Tumor-Type and Tumor-Stage splits separate patients, not merely fields of view.
    The split descriptions in Section 2.1 refer to patient groups (adenocarcinoma versus mucinous, stage 3 versus stage 4) but do not explicitly state that no patient contributes to both training and test sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PhenoBench: A Comprehensive Benchmark for Cell Phenotyping." pith.science (2026). https://pith.science/paper/BG4ONOPX

@misc{pith2026250703532,
  author       = {Pith},
  title        = {Pith review of: PhenoBench: A Comprehensive Benchmark for Cell Phenotyping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BG4ONOPX}},
  note         = {Machine review of arXiv:2507.03532}
}
read the original abstract

Digital pathology has seen the advent of a wealth of foundational models (FM), yet to date their performance on cell phenotyping has not been benchmarked in a unified manner. We therefore propose PhenoBench: A comprehensive benchmark for cell phenotyping on Hematoxylin and Eosin (H&E) stained histopathology images. We provide both PhenoCell, a new H&E dataset featuring 14 granular cell types identified by using multiplexed imaging, and ready-to-use fine-tuning and benchmarking code that allows the systematic evaluation of multiple prominent pathology FMs in terms of dense cell phenotype predictions in different generalization scenarios. We perform extensive benchmarking of existing FMs, providing insights into their generalization behavior under technical vs. medical domain shifts. Furthermore, while FMs achieve macro F1 scores > 0.70 on previously established benchmarks such as Lizard and PanNuke, on PhenoCell, we observe scores as low as 0.20. This indicates a much more challenging task not captured by previous benchmarks, establishing PhenoCell as a prime asset for future benchmarking of FMs and supervised models alike. Code and data are available on GitHub.

Figures

Figures reproduced from arXiv: 2507.03532 by the authors.

Figure 1
Figure 1. Overview of the PhenoCell dataset and FM benchmarking: (a) Example image from PhenoCell, a zoomed-in visualization of a tile and its segmentation mask. Along the pixel-wise panoptic annotations, information such as multiplexed immunofluores￾cence stainings, granular labels, or patient data are provided. (b) Illustration of the benchmarking pipeline for evaluation of cell phenotyping capabilities of FM. plications in… view at source ↗
Figure 2
Figure 2. Overview of Benchmarking results: (a) F1 scores of all FMs and HoVer-NeXt on the four datasets. Results from the linear probe are in white, UNetR are in dataset color. (b) F1 Scores per cell type for PhenoCell dataset. (c) Qualitative example of our predictions and ground truth for a sample with a rare cell type (DC). The choice of decoder has a varying effect between models and datasets (Fig. 2a). The greatest perf… view at source ↗
Figure 3
Figure 3. Evaluation of FMs and HoVer-NeXt on medical domain splits. (a) Lizard split on medical centers for domain generalization. Here, the FMs outperform HoVer-NeXt. (b) PhenoCell split on cancer stages 3 and 4. (c) PhenoCell split on adenocarcinoma and mucinous adenocarcinoma. Top Mean performance on base split vs. domain split. Bottom Performance by cell type. whereas foundation models decrease on average by 0.06 F1, wit… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Performance of FMs with respect to the number of trainable parameters on the left and the number of WSIs on the right. (b) Performance of FMs with UNetR head and HoVer-NeXt with a reduced finetuning training set size on Lizard and PhenoCell. settings with limited l…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy

    cs.CV 2026-06 conditional novelty 6.0 of 10

    An Atlas-foundation-model system for multi-cancer H&E tissue and cell profiling matches pathologist H&E accuracy against IHC-informed consensus and generalizes across 1,500+ cases.

  2. Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    Atlas H&E-TME is a new AI system for cell-level tissue profiling on H&E slides that matches pathologist performance when validated against an IHC-informed consensus and a large multi-cancer H&E annotation set.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Medical Imaging with Deep Learning (2024)

    Baumann, E., Dislich, B., Rumberger, J.L., Nagtegaal, I.D., Martinez, M.R., Zlobec, I.: Hover-next: A fast nuclei segmentation and classification pipeline for next generation histopathology. In: Medical Imaging with Deep Learning (2024)

    work page 2024

  2. [2]

    Nature Medicine 30(3), 850–862 (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F.K., Jaume, G., Song, A.H., Chen, B.,Zhang,A.,Shao,D.,Shaban,M.,Williams,M.,Oldenburg,L.,Weishaupt,L.L., Wang,J.J.,Vaidya,A.,Le,L.P.,Gerber,G.,Sahai,S.,Williams,W.,Mahmood,F.: Towards a general-purpose foundation model for computational pathology. Nature Medicine 30(3), 850–862 (2024)

    work page 2024

  3. [3]

    In: IEEE/CVF conference on computer vision and pattern recognition

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)

    work page 2022

  4. [5]

    arXiv preprint arXiv:2401.04079 (2024)

    Dippel, J., Feulner, B., Winterhoff, T., Milbich, T., Tietz, S., Schallenberg, S., Dernbach, G., Kunft, A., Heinke, S., Eich, M.L., et al.: Rudolfv: a foundation model by pathologists for pathologists. arXiv preprint arXiv:2401.04079 (2024)

    arXiv 2024

  5. [6]

    Journal of Proteome Research14(6), 2707–2713 (2015)

    Edwards, N.J., Oberti, M., Thangudu, R.R., Cai, S., McGarvey, P.B., Jacob, S., Madhavan, S., Ketchum, K.A.: The CPTAC Data Portal: A Resource for Cancer Proteomics Research. Journal of Proteome Research14(6), 2707–2713 (2015)

    work page 2015

  6. [7]

    arXiv preprint arXiv:2409.09173 (2024)

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

    arXiv 2024

  7. [8]

    Ad- vances in Neural Information Processing Systems37, 71855–71867 (2025) 10 C

    Franzen, J., Winklmayr, C., Guarino, V.E., Karg, C., Yu, X., Koreuber, N., Al- brecht, J., Bischoff, P., Kainmueller, D.: Arctique: An artificial histopathologi- cal dataset unifying realism and controllability for uncertainty quantification. Ad- vances in Neural Information Processing Systems37, 71855–71867 (2025) 10 C. Winklmayr, J. Lüscher, N. Koreuber...

    work page 2025

  8. [9]

    arXiv preprint arXiv:2003.10778 (2020)

    Gamper, J., Koohbanani, N.A., Benes, K., Graham, S., Jahanifar, M., Khurram, S.A., Azam, A., Hewitt, K., Rajpoot, N.: Pannuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778 (2020)

    arXiv 2020

Show all 27 references
  1. [10]

    In: IEEE/CVF international con- ference on computer vision

    Graham, S., Jahanifar, M., Azam, A., Nimir, M., Tsang, Y.W., Dodd, K., Hero, E., Sahota, H., Tank, A., Benes, K., et al.: Lizard: A large-scale dataset for colonic nuclear instance segmentation and classification. In: IEEE/CVF international con- ference on computer vision. pp....

    2021

  2. [11]

    Medical image analy- sis 92, 103047 (2024)

    Graham, S., Vu, Q.D., Jahanifar, M., Weigert, M., Schmidt, U., Zhang, W., Zhang, J., Yang, S., Xiang, J., Wang, X., et al.: Conic challenge: Pushing the frontiers of nuclear detection, segmentation, classification and counting. Medical image analy- sis 92, 103047 (2024)

    2024

  3. [12]

    The Genotype-Tissue Expression (GTEx) pilot analysis: Multitissue gene regulation in humans

    GTEx Consortium: Human genomics. The Genotype-Tissue Expression (GTEx) pilot analysis: Multitissue gene regulation in humans. Science (New York, N.Y.) 348(6235), 648–660 (2015)

    2015

  4. [13]

    In: IEEE/CVF winter conference on applications of computer vision

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: IEEE/CVF winter conference on applications of computer vision. pp. 574–584 (2022)

    2022

  5. [14]

    NeurIps (2023)

    Ikezogwo, W.O., Seyfioglu, M.S., Ghezloo, F., Geva, D.S.C., Mohammed, F.S., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1M: One Million Image-Text Pairs for Histopathology. NeurIps (2023)

    2023

  6. [15]

    NCI: The Cancer Genome Atlas Program (TCGA) - NCI (05/13/2022 - 08:00)

    2022

  7. [16]

    arXiv preprint arXiv:2408.15823 (2024)

    Neidlinger, P., El Nahhas, O.S., Muti, H.S., Lenz, T., Hoffmeister, M., Brenner, H., van Treeck, M., Langer, R., Dislich, B., Behrens, H.M., et al.: Benchmark- ing foundation models as feature extractors for weakly-supervised computational pathology. arXiv preprint arXiv:2408....

    2024 arXiv

  8. [17]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P....

    2023

  9. [18]

    In: 2022 IEEE International Symposium on Biomedical Imaging Challenges (ISBIC)

    Rumberger, J.L., Baumann, E., Hirsch, P., Janowczyk, A., Zlobec, I., Kainmueller, D.: Panoptic segmentation with highly imbalanced semantic labels. In: 2022 IEEE International Symposium on Biomedical Imaging Challenges (ISBIC). pp. 1–4. IEEE (2022)

    2022

  10. [19]

    Nature methods 13(6), 493–496 (2016)

    Samusik, N., Good, Z., Spitzer, M.H., Davis, K.L., Nolan, G.P.: Automated map- ping of phenotype space with single-cell data. Nature methods 13(6), 493–496 (2016)

    2016

  11. [20]

    Cell 182(5), 1341–1359 (2020)

    Schürch, C.M., Bhate, S.S., Barlow, G.L., Phillips, D.J., Noti, L., Zlobec, I., Chu, P., Black, S., Demeter, J., McIlwain, D.R., et al.: Coordinated cellular neighbor- hoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell 182(5), 1341–1359 (2020)

    2020

  12. [21]

    In: AAAI Conference on Artificial Intelligence

    Sun, Y., Zhu, C., Zheng, S., Zhang, K., Sun, L., Shui, Z., Zhang, Y., Li, H., Yang, L.: PathAsst: A Generative Foundation AI Assistant towards Artificial General Intelligence of Pathology. In: AAAI Conference on Artificial Intelligence. vol. 38, pp. 5034–5042 (2024)

    2024

  13. [22]

    Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mo- hammed, O.K., Singhal, S., Som, S., Wei, F.: Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks (2022) PhenoBench: A Comprehensive Benchmark for Cell Phenotyping 11

    2022

  14. [23]

    Nature pp

    Xiang, J., Wang, X., Zhang, X., Xi, Y., Eweje, F., Chen, Y., Li, Y., Bergstrom, C., Gopaulchan, M., Kim, T., Yu, K.H., Willens, S., Olguin, F.M., Nirschl, J.J., Neal, J., Diehn, M., Yang, S., Li, R.: A vision–language foundation model for precision oncology. Nature pp. 1–10 (2025)

    2025

  15. [24]

    Nature630(8015), 181–188 (2024)

    Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Naumann, T., Wong, C., Gero, Z., González, J., Gu, Y., Xu, Y., Wei, M., Wang, W., Ma, S., Wei, F., Yang, J., Li, C., Gao, J., Rosemon, J., Bower, T., Lee, S., Weerasinghe, R., Wright, B.J., Robicsek, A., Piening, B., Bifulco,...

    2024

  16. [25]

    arXiv preprint arXiv:2411.02796 (2024)

    Xu, Z., Gupta, R., Cheng, W., Shen, A., Shen, J., Talwalkar, A., Khodak, M.: Specialized foundation models struggle to beat supervised baselines. arXiv preprint arXiv:2411.02796 (2024)

    2024 arXiv

  17. [26]

    Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. 2022 (2022)

    2022

  18. [27]

    In: International Conference on Learning Rep- resentations (2022)

    Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image BERT Pre-training with Online Tokenizer. In: International Conference on Learning Rep- resentations (2022)

    2022

  19. [28]

    Zimmermann, E., Vorontsov, E., Viret, J., Casson, A., Zelechowski, M., Shaikovski, G., Tenenholtz, N., Hall, J., Klimstra, D., Yousfi, R., Fuchs, T., Fusi, N., Liu, S., Severson, K.: Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology (2024)

    2024

Pith tools

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