Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Towards Comprehensive Cellular Characterisation of H&E slides

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that HistoPLUS, a compact model combining the CellViT architecture with a distilled pathology encoder, outperforms larger state-of-the-art models in H&E cell detection and classification while using five times fewer paramet

desk verdict Useful new dataset and benchmark, but the headline SOTA claim doesn't survive a fair baseline comparison; the real story is efficiency-parity. read the letter →

arxiv 2508.09926 v3 pith:NRKIB6PX submitted 2025-08-13 cs.CV q-bio.QM

Benjamin Adjadj , Pierre-Antoine Bannier , Guillaume Horent , Sebastien Mandela , Aurore Lyon , Kathryn Schutte , Ulysse Marteau , Valentin Gaury
show 8 more authors
Laura Dumont Thomas Mathieu MOSAIC consortium Reda Belbahri Benoît Schmauch Eric Durand Katharina Von Loga Lucie Gillet
This is my paper
classification cs.CVq-bio.QM
keywords histopathologyH&Ecelldetectionsegmentationclassificationpathologyfoundationmodeltumormicroenvironmentactivelearning
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

HistoPLUS is presented as a single model that detects, segments, and classifies 13 cell types on H&E whole-slide images, including cell types largely missing from public datasets. The paper claims that on external validation it beats the CellViT model with the SAM-H encoder, the authors' state-of-the-art baseline, by 5.2% in detection quality and 23.7% in overall F1, while using about five times fewer parameters. This matters because routine H&E slides could then be used to quantify immune and stromal populations—neutrophils, eosinophils, plasmocytes, smooth muscle and endothelial cells, mitotic figures, and apoptotic bodies—normally studied with special stains or not at all. The recipe combines a curated active-learning pan-cancer dataset of 108,722 nuclei with a compact pathology foundation encoder inside the CellViT architecture.

What carries the argument

The machine carrying the argument is the CellViT architecture—a vision transformer encoder with three decoder branches producing nuclei masks, horizontal-vertical distance maps, and per-nucleus type maps—combined with H0-mini, a self-supervised pathology encoder distilled from a much larger foundation model. Ground truth for both training and validation is produced by NuClick, a point-to-contour segmentation network that expands pathologist centroid clicks into nucleus boundaries, so a human only labels a point and a class. HistoTRAIN, the resulting dataset, is deliberately enriched with rare phenotypes through clustering and active learning, which gives the model exposure to understudied ce

What would settle it

Take a random sample of HistoVAL tiles, have two or three pathologists manually trace nuclear boundaries instead of clicking centroids, and recompute detection quality, segmentation quality, and per-class F1 against the manual contours. If the manual-contour metrics differ materially from the NuClick-based metrics, or if the classification gap between HistoPLUS and CellViT SAM-H narrows to non-significance, the reported superiority is an artifact of the NuClick ground truth.

Watch

Extended reading notes

Core claim

The core claim is that domain-specific self-supervised pretraining, not model scale, is what lifts cell classification performance on H&E. A compact distilled pathology encoder, H0-mini (86M parameters), inside the CellViT architecture matches or beats huge encoders such as UNI2 and Virchow2 on external validation, while exceeding the general-purpose SAM-H encoder on rare and understudied classes. Trained on HistoTRAIN and validated on the consensus-annotated HistoVAL, this configuration—released as HistoPLUS—generalizes to two cancer indications unseen at training and brings significant F1 gains on 8 of 13 cell types.

Load-bearing premise

The load-bearing premise is that NuClick, which turns a single centroid click into a nucleus outline, produces segmentations accurate enough to serve as true nuclear boundaries; if those contours are systematically wrong, every detection, segmentation, and classification score—on both training and validation—is measured against the wrong ground truth.

Editorial extensions

If this is right

  • Seven cell types that public datasets mostly omit become measurable on routine H&E slides, unlocking new windows on immune and stromal tumor-microenvironment composition.
  • A compact model can match or beat encoders with several times more parameters, so high-throughput cellular profiling does not demand huge GPU infrastructure.
  • External validation on breast and ovarian cancer suggests the model transfers to indications not seen in training, at least for detection and key classes such as lymphocytes and cancer cells.
  • Segmentation quality stays roughly flat across encoder sizes, indicating that the benefit of pathology-specific pretraining concentrates in classification rather than boundary delineation.
  • With model weights and inference code released, independent groups can test cellular composition and spatial relationships against survival or treatment-response endpoints.

Reading between the lines

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

  • My inference: if NuClick boundary bias is shared between training and validation ground truth, the reported 23.7% F1 gap could change under manual-contour scoring; an independent manual delineation rebenchmark would settle the true effect size.
  • My inference: the recipe of a distilled pathology encoder plus active-learning enriched rare-class data is likely transferable to other dense cell-typing tasks, such as immunohistochemistry or multiplex immunofluorescence, where rare classes are the bottleneck.
  • My inference: the two unseen indications are still external cohorts from the same data ecosystem; a prospective cohort with different scanners and staining protocols would test real-world deployment behavior, which the paper itself flags as future work.
  • My inference: combined with tissue-level spatial models, HistoPLUS could yield spatial biomarkers—such as juxtatumoral neutrophil density or plasmocyte distribution near tumor boundaries—that density-only analyses miss; the paper gestures at this but does not validate it.
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 / 4 minor

Summary. The paper presents HistoPLUS, a cell detection, segmentation, and classification model built on the CellViT architecture with an H0-mini pathology foundation-model encoder. It is trained on HistoTRAIN, a newly introduced pan-cancer dataset of 108,722 nuclei covering 13 cell types, and validated on HistoVAL, a consensus-based external set built from multiple pathologist annotations. The authors claim that HistoPLUS outperforms current state-of-the-art models in detection quality by 5.2% and overall F1 by 23.7% while using 5x fewer parameters, and that it transfers robustly to two unseen cancer indications. The paper also contributes an active-learning annotation pipeline, a systematic comparison of pathology foundation models as CellViT encoders, and released model weights and inference code.

Significance. If the central claims are properly supported, this would be a useful contribution: the pan-cancer dataset with rare cell-type enrichment addresses a real gap in public resources, the active-learning and consensus-validation pipelines are methodologically interesting, and the comparison of multiple PFM encoders within a fixed architecture is a valuable benchmark. The release of weights and code, deterministic training with a fixed seed, and explicit training details are strengths that support reproducibility. However, the headline SOTA claim is currently based on a selective baseline comparison and a validation set that was also used for model selection; the segmentation ground truth is generated by a trained algorithm rather than independent human contours. These issues are central to the paper's claims and require substantial revision.

major comments (3)
  1. [Abstract / §2.3 / Fig. 4b / Supplementary Tables 3–4] The headline claim that HistoPLUS outperforms current state-of-the-art models by 5.2% in detection quality and 23.7% in F1 is computed against CellViT with SAM-H only. On the paper's own external validation, SAM-H is the weakest large encoder: DQ 0.716 vs 0.753 for H0-mini, UNI2, and Virchow2 (Supplementary Table 3). The macro-average F1 computed from Supplementary Table 4 is approximately 0.391 for H0-mini, 0.396 for UNI2, 0.367 for Virchow2, and 0.313 for SAM-H. Thus the 5.2% DQ gain and most of the 23.7% F1 gain are artifacts of choosing SAM-H as the baseline; against UNI2/Virchow2 the DQ advantage disappears and the F1 advantage is negligible or reversed. The claim should be reframed as efficiency-parity with current large PFMs, or supported by a comparison against the strongest available baselines on a held-out set.
  2. [§2.2 and Methods, 'Cell boundary extension with Nuclick'] Both HistoTRAIN and HistoVAL ground-truth contours are generated by NuClick, a trained CNN, from pathologist point annotations. The consensus framework in §2.2 also uses NuClick segmentations for matching and for final contours. Consequently, the segmentation quality (SQ) and the IoU-based detection-quality matching measure agreement with the label-generation algorithm rather than with independent human boundary delineations. Because HistoPLUS is trained on NuClick-derived contours and evaluated on the same kind of contours, the absolute segmentation figures are partly circular. This does not necessarily invalidate relative comparisons between encoders, since all models see the same ground truth, but the paper should present the numbers as 'NuClick-consensus' segmentation quality and, ideally, anchor them with a small set of independent manual delineations.
  3. [§2.3, Supplementary Tables 3–4] The conclusion that ViT-Huge encoders provide only marginal gains over H0-mini is based on informal comparison of overlapping confidence intervals. For example, DQ is identical (0.753) for H0-mini, UNI2, and Virchow2, but UNI2 has a higher macro-F1 (≈0.396 vs ≈0.391) and higher per-class F1 on several classes such as cancer cells (0.603 vs 0.583) and plasmocytes (0.507 vs 0.483). Without paired significance tests or an equivalence test on the relevant pairwise differences, the claim that HistoPLUS is the best accuracy-efficiency trade-off is not statistically established. In addition, §2.3 states that external results 'motivated the selection' of HistoPLUS; using HistoVAL for model selection and then reporting it as the test bed for the headline comparison biases the evaluation.
minor comments (4)
  1. [Table 2 caption] 'Zero-short generalizability' should be 'Zero-shot generalizability'.
  2. [§2.4, Table 2] There is a missing closing bracket in '0.801 [0.796 - 0.807 and 0.803 [0.795 - 0.810]' — the first interval should be '[0.796 - 0.807]'.
  3. [Throughout] The spelling of 'NuClick' is inconsistent ('Nuclick' vs 'NuClick'); please standardize. Also, 'HistoV AL' and 'HistoTRAIN' appear with irregular spacing in several places.
  4. [Methods, 'Vision Transformer'] The equation '(𝑥 𝑝) 𝑝∈R𝑁×(𝑃×𝐶) ' has formatting issues and should be typeset cleanly. Minor typo: 'non overlapping areas' should be 'non-overlapping areas'.

Circularity Check

2 steps flagged · score 5.0 of 10

HistoPLUS's headline numbers are partly circular: the segmentation/detection ground truth is NuClick-generated for both training and validation, and HistoVAL was used to select the model and then reused as the headline test set.

  1. fitted input called prediction [Methods, 'Cell boundary extension with Nuclick'; Results §2.2 (HistoVAL construction); Supplementary Figure 1]
    "Nuclick segmentations are then used as ground truth annotations for model training. ... The final consensus segmentations are generated by applying NuClick to the computed consensus centroids, yielding high-quality nuclear boundaries paired with robust cell type classifications."

    Both the training targets and the external-validation ground-truth contours are NuClick outputs. DQ/SQ are computed by IoU against these NuClick masks, so the segmentation/detection metrics measure agreement with the same learned label generator that the model was trained to imitate, not with independent manual boundaries. The paper presents these as 'high-quality nuclear boundaries', but the ground truth is by construction a model prediction. This partly reduces the reported segmentation/detection quality to a fit-to-NuClick score. It does not force the relative ranking because all baselines share the same labels, but it undermines the absolute quality interpretation and weakens the external validity of the reported DQ/SQ numbers.

  2. other [Section 2.3 (final paragraph) and Figure 4b; external validation described in Methods]
    "Taken together, these results motivated the selection of HistoPLUS, a compact yet high-performing configuration combining the CellViT [40] architecture with the H0-mini [36] backbone, as our reference model. ... b) Head-to-head comparison between CellViT with SAM-H (state-of-the-art) and our model HistoPLUS – CellViT with the H0-mini encoder – on our external test set."

    HistoVAL was first used in Section 2.3 to compare all encoders and to motivate choosing H0-mini as the HistoPLUS backbone; the same HistoVAL is then used in Figure 4b as the 'external test set' for the headline 5.2%/23.7% comparison. The reported external performance is therefore not an independent holdout result: the model configuration was selected on this set, so the headline numbers are partly a product of the selection loop. This is a form of fitting the model choice to the test set and then calling the test-set result a prediction.

full rationale

The core derivation of HistoPLUS as a CellViT+H0-mini combination is not circular in the equation sense: it is an architecture/checkpoint choice, and the experiments comparing encoders are real. The two circular/leaky steps are: (1) the segmentation/detection ground truth is NuClick-generated for both training and validation, so SQ/DQ are partly self-comparisons against a learned label generator rather than independent human contours; and (2) HistoVAL was used to motivate the selection of the H0-mini backbone and then reused as the test set for the headline comparison, so the headline numbers are partly a selection result. These do not force the classification-F1 comparisons, since class labels come from pathologist majority voting, so the paper retains independent content. I also note as a non-circular benchmarking concern that the 5.2% DQ claim is computed against SAM-H, which the paper's own Supplementary Table 3 ranks below UNI2/Virchow2/H0-mini on DQ (0.716 vs 0.753); this affects the SOTA framing but is not a circularity. No load-bearing self-citation or imported uniqueness theorem was found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or ontological entities. The central claim rests on trained model parameters (NuClick, H0-mini, CellViT heads, presence MLPs), on hand-chosen dataset construction thresholds, and on the domain assumption that algorithm-derived segmentations and pathologist point votes constitute valid ground truth.

free parameters (6)
  • NuClick trained weights = trained on ConSeP and PCNS (validation Dice 0.874)
    Used to generate all nucleus segmentations in HistoTRAIN and HistoVAL; parameters are fitted to external datasets.
  • H0-mini encoder weights = pretrained, 86M parameters
    Frozen feature extractor distilled from H-Optimus-0; fitted to large pathology corpora by the original authors.
  • CellViT decoder and supervision head weights = trained on HistoTRAIN, 150 epochs, seed 111
    Fitted to the 108,722 annotated nuclei; this is the central predictive model.
  • Active-learning MLP presence predictors = trained on TCGA COAD/LUAD/LUSC tiles (636 to 1782 tiles)
    Used to select tiles for annotation and to construct HistoVAL; their predictions shape dataset composition.
  • K-means cluster count K = 20
    Hand-chosen based on literature [47]; controls tile diversity sampling.
  • Consensus matching IoU threshold = 0.4
    Hand-chosen; determines which nuclei are considered corresponding across pathologists in HistoVAL.
assumptions (4)
  • domain assumption NuClick model outputs are an accurate proxy for true nuclear boundaries
    Both HistoTRAIN and HistoVAL segmentation labels are derived by running NuClick on pathologist point annotations; if NuClick contours are biased, all segmentation metrics are biased. Methods: 'Cell boundary extension with Nuclick' and Section 2.2.
  • domain assumption Pathologist point annotations and majority voting yield correct cell-type ground truth
    Cell classes are assigned by 2-3 pathologists with majority vote; this assumes majority opinion is ground truth, despite acknowledged inter-annotator variability. Section 2.2.
  • domain assumption H0-mini pretrained features transfer to H&E cell morphology across cancer types
    The model relies on H0-mini, a distilled pathology foundation model pretrained on unlabeled histology; assumes its features are informative for cell classification in the six training indications and two unseen indications. Section 2.3.
  • domain assumption External validation set provides an unbiased estimate of generalization
    The paper reports HistoVAL as an external validation, but the final model was selected based on performance on this set, so this assumption is partially violated. Section 2.3: 'These results motivated the selection of HistoPLUS'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Comprehensive Cellular Characterisation of H&E slides." pith.science (2026). https://pith.science/paper/NRKIB6PX

@misc{pith2026250809926,
  author       = {Pith},
  title        = {Pith review of: Towards Comprehensive Cellular Characterisation of H&E slides},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NRKIB6PX}},
  note         = {Machine review of arXiv:2508.09926}
}
read the original abstract

Cell detection, segmentation and classification are essential for analyzing tumor microenvironments (TME) on hematoxylin and eosin (H&E) slides. Existing methods suffer from poor performance on understudied cell types (rare or not present in public datasets) and limited cross-domain generalization. To address these shortcomings, we introduce HistoPLUS, a state-of-the-art model for cell analysis, trained on a novel curated pan-cancer dataset of 108,722 nuclei covering 13 cell types. In external validation across 4 independent cohorts, HistoPLUS outperforms current state-of-the-art models in detection quality by 5.2% and overall F1 classification score by 23.7%, while using 5x fewer parameters. Notably, HistoPLUS unlocks the study of 7 understudied cell types and brings significant improvements on 8 of 13 cell types. Moreover, we show that HistoPLUS robustly transfers to two oncology indications unseen during training. To support broader TME biomarker research, we release the model weights and inference code at https://github.com/owkin/histoplus/.

Figures

Figures reproduced from arXiv: 2508.09926 by the authors.

Figure 1
Figure 1. Overview of our training dataset. a) Distribution of annotated nuclei across cell types and cancer indications. Barplot showing the number of annotated nuclei per cell type across six cancer types. The active learning pipeline enhanced the enrichment of our dataset in understudied cell types on H&E. b) Candidate whole-slide images (WSIs) are divided into tiles, and features are extracted from each tile using Phikon,… view at source ↗
Figure 2
Figure 2. Robust external validation sets. Example of consensus annotations derived from regions annotated independently by 3 expert-pathologists. Disagreement may arise from distinct nuclei chosen to be annotated vs. skipped [1], distinct cell types being attributed [2], or a mix of both of these reasons [3], [4]. To compare equal-sized encoders, we evaluated PFMs built on the vision transformer [42] (ViT) base architec￾ture… view at source ↗
Figure 3
Figure 3. Evaluation of pathology foundation models with the CellViT architecture. a) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: HistoPLUS outperforms the current state-of-the-art in detection and classification while reducing [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

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. Full citation record

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

    cs.CV 2026-06 unverdicted novelty 6.0 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

57 extracted references · 55 canonical work pages · cited by 1 Pith paper

  1. [1]

    J. A. Diao et al. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes. Nat. Commun., 12:1613, 2021

    work page 2021

  2. [2]

    Abel et al

    J. Abel et al. Ai powered quantification of nuclear morphology in cancers enables prediction of genome instability and prognosis. Npj Precis. Oncol., 8:1–14, 2024

    work page 2024

  3. [3]

    L. He, L. R. Long, S. Antani, and G. R. Thoma. Histology image analysis for carcinoma detection and grading. Comput. Methods Programs Biomed., 107:538–556, 2012

    work page 2012

  4. [4]

    E. G. Fischer. Nuclear morphology and the biology of cancer cells. Acta Cytol., 64:511–519, 2020

    work page 2020

  5. [5]

    D. Zink, A. H. Fischer, and J. A. Nickerson. Nuclear structure in cancer cells. Nat. Rev. Cancer, 4:677–687, 2004

    work page 2004

  6. [6]

    K. E. Blise, S. Sivagnanam, G. L. Banik, L. M. Coussens, and J. Goecks. Single-cell spatial architectures associated with clinical outcome in head and neck squamous cell carcinoma. NPJ Precis. Oncol., 6:10, 2022

    work page 2022

  7. [7]

    Arora et al

    R. Arora et al. Spatial transcriptomics reveals distinct and conserved tumor core and edge architectures that predict survival and targeted therapy response. Nat. Commun., 14:5029, 2023

    work page 2023

  8. [8]

    Patkar et al

    S. Patkar et al. Predicting the tumor microenvironment composition and immunotherapy response in non-small cell lung cancer from digital histopathology images. Npj Precis. Oncol., 8:1–15, 2024

    work page 2024

Show all 57 references
  1. [9]

    H. L. Williams et al. The current landscape of spatial biomarkers for prediction of response to immune checkpoint inhibition. Npj Precis. Oncol., 8:1–18, 2024

    2024

  2. [10]

    M. T. Masucci, M. Minopoli, and M. V. Carriero. Tumor associated neutrophils. their role in tumorige- nesis, metastasis, prognosis and therapy. Front. Oncol., 9, 2019

    2019

  3. [11]

    B. Feng, J. Wu, B. Shen, F. Jiang, and J. Feng. Cancer-associated fibroblasts and resistance to anticancer therapies: status, mechanisms, and countermeasures. Cancer Cell Int., 22:166, 2022

    2022

  4. [12]

    Z. Zhao, T. Li, L. Sun, Y. Yuan, and Y. Zhu. Potential mechanisms of cancer-associated fibroblasts in therapeutic resistance. Biomed. Pharmacother., 166:115425, 2023

    2023

  5. [13]

    Long et al

    X. Long et al. Cancer-associated fibroblasts promote cisplatin resistance in bladder cancer cells by increasing igf-1/erb/bcl-2 signalling. Cell Death Dis., 10:1–16, 2019

    2019

  6. [14]

    H. R. Ali et al. Lymphocyte density determined by computational pathology validated as a predictor of response to neoadjuvant chemotherapy in breast cancer: secondary analysis of the artemis trial. Ann. Oncol. Off. J. Eur. Soc. Med. Oncol., 28:1832–1835, 2017

    2017

  7. [15]

    Combination of mk3475 and metronomic cyclophosphamide in patients with advanced sarcomas : Multicentre phase ii trial

    Institut Bergoni ´e. Combination of mk3475 and metronomic cyclophosphamide in patients with advanced sarcomas : Multicentre phase ii trial. https://clinicaltrials.gov/study/NCT02406781, 2021

    2021

  8. [16]

    A. H. Fischer, K. A. Jacobson, J. Rose, and R. Zeller. Hematoxylin and eosin staining of tissue and cell sections. CSH Protoc., 2008:pdb.prot4986, 2008

    2008

  9. [17]

    Kumar et al

    N. Kumar et al. A multi-organ nucleus segmentation challenge. IEEE Trans. Med. Imaging, 39:1380– 1391, 2020

    2020

  10. [18]

    Mahbod et al

    A. Mahbod et al. Nuinsseg: A fully annotated dataset for nuclei instance segmentation in h&e-stained histological images, 2023

    2023

  11. [19]

    Q. D. Vu et al. Methods for segmentation and classification of digital microscopy tissue images, 2018

    2018

  12. [20]

    Amgad et al

    M. Amgad et al. Nucls: A scalable crowdsourcing, deep learning approach and dataset for nucleus classification, localization and segmentation. GigaScience, 11:giac037, 2022

    2022

  13. [21]

    Gamper et al

    J. Gamper et al. Pannuke dataset extension, insights and baselines, 2020

    2020

  14. [22]

    Graham et al

    S. Graham et al. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Med. Image Anal., 58:101563, 2019

    2019

  15. [23]

    N. P. Jack, W. Thomas, L. Marick, and R. Fabien. Segmentation of nuclei in histopathology images by deep regression of the distance map, 2018

    2018

  16. [24]

    Graham et al

    S. Graham et al. Lizard: A large-scale dataset for colonic nuclear instance segmentation and classifica- tion, 2021

    2021

  17. [25]

    Naylor, M

    P. Naylor, M. Lae, F. Reyal, and T. Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Trans. Med. Imaging, 38:448–459, 2019

    2019

  18. [26]

    Zhou et al

    J. Zhou et al. ibot: Image bert pre-training with online tokenizer, 2022

    2022

  19. [27]

    Oquab et al

    M. Oquab et al. Dinov2: Learning robust visual features without supervision, 2023

    2023

  20. [28]

    R. J. Chen and R. G. Krishnan. Self-supervised vision transformers learn visual concepts in histopathol- ogy, 2022

    2022

  21. [29]

    R. J. Chen et al. Towards a general-purpose foundation model for computational pathology. Nat. Med., 30:850–862, 2024

    2024

  22. [30]

    Filiot, P

    A. Filiot, P. Jacob, A. Mac Kain, and C. Saillard. Phikon-v2, a large and public feature extractor for biomarker prediction, 2024

    2024

  23. [31]

    Filiot et al

    A. Filiot et al. Scaling self-supervised learning for histopathology with masked image modeling, 2023

    2023

  24. [32]

    Saillard et al

    C. Saillard et al. H-optimus-0, 2024

    2024

  25. [33]

    https://arxiv.org/ html/2502.06750v1, 2025

    Accelerating data processing and benchmarking of ai models for pathology. https://arxiv.org/ html/2502.06750v1, 2025

    2025 arXiv

  26. [34]

    Nechaev, A

    D. Nechaev, A. Pchelnikov, and E. Ivanova. Hibou: A family of foundational vision transformers for pathology, 2024

    2024

  27. [35]

    Vadori, A

    V. Vadori, A. Peruffo, J.-M. Gra¨ıc, L. Finos, and E. Grisan. Mind the gap: Evaluating patch embeddings from general-purpose and histopathology foundation models for cell segmentation and classification, 2025

    2025

  28. [36]

    Filiot et al

    A. Filiot et al. Distilling foundation models for robust and efficient models in digital pathology, 2025

    2025

  29. [37]

    N. A. Koohbanani, M. Jahanifar, N. Z. Tajadin, and N. Rajpoot. Nuclick: A deep learning framework for interactive segmentation of microscopy images, 2020

    2020

  30. [38]

    C. Kang, C. Lee, H. Song, M. Ma, and S. Pereira. Variability matters : Evaluating inter-rater variability in histopathology for robust cell detection, 2022

    2022

  31. [39]

    Mosaic: Intra-tumoral heterogeneity characterization through large-scale spatial and cell-resolved multi-omics profiling

    MOSAIC Consortium and Caroline Hoffmann. Mosaic: Intra-tumoral heterogeneity characterization through large-scale spatial and cell-resolved multi-omics profiling. bioRxiv, 2025

    2025

  32. [40]

    H ¨orst et al

    F. H ¨orst et al. Cellvit: Vision transformers for precise cell segmentation and classification. Med. Image Anal., 94:103143, 2024

    2024

  33. [41]

    Kirillov et al

    A. Kirillov et al. Segment anything, 2023

    2023

  34. [42]

    Dosovitskiy et al

    A. Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    2021

  35. [43]

    Zimmermann et al

    E. Zimmermann et al. Virchow2: Scaling self-supervised mixed magnification models in pathology, 2024

    2024

  36. [44]

    Mar ´ee et al

    R. Mar ´ee et al. Collaborative analysis of multi-gigapixel imaging data using cytomine. Bioinformatics, 32:1395–1401, 2016

    2016

  37. [45]

    Dehaene, A

    O. Dehaene, A. Camara, O. Moindrot, A. de Lavergne, and P. Courtiol. Self-supervision closes the gap between weak and strong supervision in histology, 2020

    2020

  38. [46]

    Filiot, R

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

    2023

  39. [47]

    A. L. Meirelles, T. Kurc, J. Saltz, and G. Teodoro. Effective active learning in digital pathology: A case study in tumor infiltrating lymphocytes. Comput. Methods Programs Biomed., 220:106828, 2022

    2022

  40. [48]

    Y. Gal, R. Islam, and Z. Ghahramani. Deep bayesian active learning with image data, 2017

    2017

  41. [49]

    Pan-cancer-nuclei-seg

    The Cancer Imaging Archive (TCIA). Pan-cancer-nuclei-seg. https://www. cancerimagingarchive.net/analysis-result/pan-cancer-nuclei-seg/

  42. [50]

    Vaswani et al

    A. Vaswani et al. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

    2017

  43. [51]

    J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization, 2016

    2016

  44. [52]

    X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer. Scaling vision transformers. In2022 IEEECVF Conf. Comput. Vis. Pattern Recognit. CVPR, pages 1204–1213, 2022

    2022

  45. [53]

    Hatamizadeh et al

    A. Hatamizadeh et al. Unetr: Transformers for 3d medical image segmentation, 2021

    2021

  46. [54]

    H. W. Kuhn. The hungarian method for the assignment problem. Nav.Res. Logist. Q., 2:83–97, 1955

    1955

  47. [55]

    J. Munkres. Algorithms for the assignment and transportation problems. SIAM Journal on Applied Mathematics, 1957

    1957

  48. [56]

    Kirillov, K

    A. Kirillov, K. He, R. Girshick, C. Rother, and P. Doll´ar. Panoptic segmentation, 2019. Supplementary Material Encoder Panoptic Quality Detection Quality Segmentation Quality SAM-B 0.599 (0.011) 0.725 (0.009) 0.817 (0.004) Phikon 0.606 (0.002) 0.731 (0.003) 0.823 (0.001) Hibo...

    2019

  49. [57]

    Supplementary Fig.2

    consensus centroids are computed from the pathologist-annotated centroids, cell types are determined by majority voting, and final contours are generated by inferring NuClick on consensus points. Supplementary Fig.2. Resulting number of nuclei per class for all indications in ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.