CoChat
Sign Up
Cited by 704AI in cancer detectionArtificial Intelligence in Healthcare and EducationRadiomics and Machine Learning in Medical Imaging

A foundation model for clinical-grade computational pathology and rare cancers detection

Open access
Vorontsov, Eugene, Bozkurt, Alican, Casson, Adam, Shaikovski, George, Zelechowski, Michal, Severson, Kristen, Zimmermann, Eric, Hall, James +25 more·2024·Nature Medicine
DOI 10.1038/s41591-024-03141-0PMC PMC11485232
Open this paper in CoChat — free account

Abstract

The analysis of histopathology images with artificial intelligence aims to enable clinical decision support systems and precision medicine. The success of such applications depends on the ability to model the diverse patterns observed in pathology images. To this end, we present Virchow, the largest foundation model for computational pathology to date. In addition to the evaluation of biomarker prediction and cell identification, we demonstrate that a large foundation model enables pan-cancer detection, achieving 0.95 specimen-level area under the (receiver operating characteristic) curve across nine common and seven rare cancers. Furthermore, we show that with less training data, the pan-cancer detector built on Virchow can achieve similar performance to tissue-specific clinical-grade models in production and outperform them on some rare variants of cancer. Virchow’s performance gains highlight the value of a foundation model and open possibilities for many high-impact applications with limited amounts of labeled training data.

AI snapshot

Written from the paper’s full text.

TL;DR

Virchow is a 632-million-parameter vision transformer foundation model for computational pathology trained on ~1.5 million H&E whole-slide images from ~100,000 patients. It enables pan-cancer detection with 0.950 specimen-level AUC across common and rare cancers and approaches or exceeds clinical-grade specialist models on several tasks with less training data.

Key findings

  • Virchow embeddings achieved the best pan-cancer detection performance across all tested cancer types: overall AUC 0.950, versus 0.940 for UNI, 0.932 for Phikon and 0.907 for CTransPath; rare-cancer AUC was 0.937.
  • At 95% sensitivity, the Virchow-based pan-cancer model reached 72.5% specificity, compared with 68.9%, 62.9% and 52.3% for UNI, Phikon and CTransPath, respectively.
  • Against clinical products, the pan-cancer model reached AUC 0.980 on prostate, 0.985 on breast and 0.971 on breast lymph node; it was below Paige Prostate (0.995, P < 0.05) and Paige Breast (0.992, P < 0.01), but better than Paige BLN for macrometastases (0.999 versus 0.994, P < 0.05).
  • Virchow embeddings ranked top in 7 of 9 biomarker prediction tasks and matched or surpassed other embeddings in 7 of 8 tile-level benchmark tasks.

Methods

  • Design: Virchow is a ViT-H/14 model with 632 million parameters trained with DINO v.2 self-supervised learning on 2 billion tiles sampled from approximately 13 billion available tissue tiles.
  • Data: training used 1,488,550 H&E WSIs from 119,629 patients across 17 tissue groups; pan-cancer evaluation used 22,932 slides from 6,142 specimens across 16 cancer types, including internal MSKCC and external consultation cases.
  • Analysis: downstream pan-cancer and biomarker tasks used a weakly supervised Agata aggregator; tile-level tasks used linear probing. Performance was assessed mainly by AUC, specificity at 95% sensitivity, F1, DeLong’s test, Cochran’s Q/McNemar tests and Holm correction.

Limitations

  • Stated: training data came from one center with limited scanner types; embeddings are tile-level at ×20 magnification (0.5 mpp), requiring an aggregation model, and slide-level models may be needed for low-data tasks such as biomarker, treatment-response or outcome prediction.
  • Stated: rare-variant and biomarker comparisons lacked statistical power; thorough stratified clinical validation, hardware/deployment considerations and possible model distillation remain necessary, and data-balancing/distillation strategies were not fully explored.
  • Evident from the text: the evaluation is retrospective and uses deidentified secondary data, with no prospective clinical deployment study reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41591-024-03141-0

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 54 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Cancer detection scales approximately logarithmically with the number of model parameters.

    “Cancer detection was found to scale approximately logarithmically with the number of model parameters (Fig. 2g, top);”
    Quote in the paperNot a substantial claim
  • Findingown data

    Virchow is trained on data from approximately 100,000 patients corresponding to approximately 1.5 million H&E stained WSIs acquired from MSKCC.

    “Virchow is trained on data from approximately 100,000 patients corresponding to approximately 1.5 million H&E stained WSIs acquired from Memorial Sloan Kettering Cancer Center (MSKCC)”
    Quote in the paperSubstantial claim
  • Findingown data

    Virchow embeddings yielded an AUC of 0.937 on rare cancers.

    “Compared to the aforementioned AUC of 0.950 overall, Virchow embeddings yielded an AUC of 0.937 on rare cancers (Fig. 2c), demonstrating generalization to rare data.”
    Quote in the paperSubstantial claim
  • Findingown data

    Performance across individual rare cancers was non-uniform, with cervical and bone cancers more challenging (AUC < 0.9) irrespective of embeddings.

    “Performance across the individual rare cancers was, however, non-uniform, with detection of cervical and bone cancers proving more challenging (AUC < 0.9) irrespective of the embeddings used (Fig. 2a,d).”
    Quote in the paperNot a substantial claim
  • Findingown data

    Virchow embeddings improved cervix detection to 0.875 AUC compared with 0.830, 0.810, 0.753 for UNI, Phikon, CTransPath.

    “Virchow embeddings improved cervix detection to 0.875 AUC compared with 0.830, 0.810 or 0.753 when using UNI, Phikon or CTransPath embeddings, respectively.”
    Quote in the paperSubstantial claim
  • May combine two claims. 42
  • Classified as probably not a substantial claim. 27
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingcitation

    Virchow is a 632 million parameter ViT model trained using the DINO v.2 algorithm.

    “Virchow, a 632 million parameter ViT model, is trained using the DINO v.2 algorithm33, a multiview student–teacher self-supervised algorithm”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Oquab, M. et al. DINOv2: Learning Robust Visual Features without Supervision. In Transactions on Machine Learning Research 2835–8856 (TMLR, 2024).

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The pan-cancer model achieved an overall AUC of 0.950 with Virchow embeddings, 0.940 with UNI, 0.932 with Phikon, and 0.907 with CTransPath, with all differences significant (P < 0.0001).

    “Overall the pan-cancer model achieved an AUC of 0.950 with Virchow embeddings, 0.940 with UNI embeddings, 0.932 with Phikon embeddings and 0.907 with CTransPath embeddings (Fig. 2b; all significantly different with P < 0.0001).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Virchow embeddings outperform or match all baseline models on all tested cancer types, including rare cancers and out-of-distribution (OOD) data.

    “Virchow embeddings outperform or match all baseline models on all tested cancer types, notably including rare cancers and out-of-distribution (OOD) data.”
    • Also stated in: § Virchow enables pan-cancer detection · plain
      “Virchow embeddings yielded the best cancer detection performance on all cancer types (Fig. 2a).”
    • Also stated in: § Virchow enables pan-cancer detection · plain
      “Virchow embeddings enable the best cancer detection performance across all cancer types, including rare cancers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    All models achieve a similar AUC on both internal and external data, suggesting that they generalize well.

    “Although AUC cannot be exactly compared across data subsets, we can observe that all models achieve a similar AUC on both internal and external data, suggesting that they generalize well as external data can be challenging because it is submitted to MSKCC for consultation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Performance scaled with the number of training tile samples, with a trend suggesting diminishing returns.

    “although performance scaled with the number of training tile samples, the trend (Fig. 2g, bottom) suggests diminishing returns.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The training data are composed of cancerous and benign tissues, collected via biopsy (63%) and resection (37%), from 17 high-level tissues.

    “The training data are composed of cancerous and benign tissues, collected via biopsy (63%) and resection (37%), from 17 high-level tissues (Fig. 1b–d).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The Virchow-based pan-cancer detection model performs nearly as well as the prostate, breast and BLN clinical specialist models while matching or outperforming them on many rare variants of cancers.

    “The Virchow-based pan-cancer detection model, trained on cancers across numerous tissues, performs nearly as well as the prostate, breast and BLN clinical specialist models (Fig. 3c) while outperforming them on many rare variants of cancers (Fig. 3d).”
    • Also stated in: § Toward clinical-grade performance · plain
      “In addition to approaching the specialist models in terms of overall AUCs, the pan-cancer model matches or outperforms these models on rare variants of cancers, as shown in Fig. 3d.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Pan-cancer detection using UNI embeddings achieved statistically similar performance (P < 0.05) for eight of nine common and five of seven rare cancer types; in all but one case, the specific AUC was lower.

    “Pan-cancer detection using UNI embeddings achieved statistically similar performance (P < 0.05) for eight of the nine common cancer types and five of the seven rare cancer types; nevertheless, in all but one case, the specific area under (the receiver operating characteristic) curve (AUC) score was lower.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Cervix, testis and H&N are tissues not seen during training, and Virchow embeddings still outperform competing models.

    “Furthermore, cervix, testis and head and neck (H&N) are tissues not seen during training, and Virchow embeddings still outperform competing models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Performance scales with the size of the foundation model and the size of the training data.

    “The comparison of pan-cancer performance based on different foundation model embeddings reveals that performance scales with the size of the foundation model and the size of the training data (Fig. 2g).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Virchow embeddings demonstrated superior performance in seven of the nine evaluated digital biomarkers, achieving AUC scores exceeding those of the nearest baseline foundation models.

    “In our comparative analysis shown in Fig. 4c, Virchow embeddings demonstrated superior performance in seven of the nine evaluated digital biomarkers, achieving AUC scores that exceeded those of the nearest baseline foundation models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Paige Prostate was trained on 66,713 blocks, Paige Breast on 44,588 specimens, BLN on 8,150 specimens, whereas pan-cancer was trained on 35,387 groups of slides total, of which 2,829 prostate, 1,626 breast, 1,441 lymph node.

    “Concretely, Paige Prostate was trained on 66,713 blocks, Paige Breast was trained on 44,588 specimens and BLN on 8150 specimens, whereas pan-cancer (using Virchow embeddings) was trained on only 35,387 groups of slides (blocks or specimens) in total, of which 2,829 are prostate, 1,626 are breast and 1,441 are lymph node.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The pan-cancer model achieves an AUC of 0.980, 0.985 and 0.971 on prostate, breast and BLN, respectively.

    “The pan-cancer model achieves an AUC of 0.980, 0.985 and 0.971 on prostate, breast and BLN, respectively.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    The training dataset of the pan-cancer model lacks the quality control and subpopulation enrichment of data and labels that are typically done for commercially available AI models.

    “This result is even more impressive when noting that the training dataset of the pan-cancer model, as a proof of concept, lacks the quality control and subpopulation enrichment of data and labels that are typically done for commercially available AI models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    There is no statistically significant difference (P < 0.05) in the other BLN comparisons or some stratified breast cancer comparisons.

    “Furthermore, there is no statistically significant difference (P < 0.05) in the other BLN comparisons or some of the stratified breast cancer comparisons (Fig. 3c).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    UNI scored in the top 1 three times and in the top 2 for all tasks, and Phikon scored among the top 2 three times.

    “The closest competing models are UNI and Phikon, with UNI scoring in the top 1 three times and in the top 2 for all tasks and Phikon scoring in among the top 2 three times.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The improvement in performance with respect to model size appears to still be in an approximately log-linear range, while sub-log-linear trends were observed as a function of training data.

    “The improvement in performance with respect to model size appears to still be in an approximately log-linear range; however, sub-log-linear trends were observed as a function of training data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Most detected precursors exhibited high-grade dysplasia, with cytologic features resembling invasive carcinoma, although some foci of low-grade dysplasia were also detected.

    “Most detected precursors exhibited high-grade dysplasia, with cytologic features resembling those of invasive carcinoma, although some foci of low-grade dysplasia were also detected in the gastroesophageal junction and skin.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In both examples, the first principal component highlighted malignant epithelium cells, and the second principal component highlighted miscellaneous and inflammatory cells.

    “In both examples, the first principal component highlighted malignant epithelium (red) cells. The second principal component, respectively, highlighted miscellaneous cells (yellow) and inflammatory (magenta) cells.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The pan-cancer model is surpassed by Paige Prostate (0.980 vs 0.995 AUC, P < 0.05) and Paige Breast (0.985 vs 0.992 AUC, P < 0.01).

    “This performance approaches that of commercial models; however, it is still surpassed by the Food and Drug Administration-approved Paige Prostate model (0.980 versus 0.995 AUC, P < 0.05) and the Paige Breast model (0.985 versus 0.992 AUC, P < 0.01).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In 11 cases (26.2%), there was more than minimal cancer within the specimen, and the negative result of the model could not be explained.

    “In 11 cases (26.2%), there was more than minimal cancer within the specimen, and the negative result of the model could not be explained.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In breast tissue, the pan-cancer model outperforms the Paige Breast model overall and especially on some rare histological variants, including adenoid cystic carcinoma, carcinoma with apocrine differentiation (P < 0.05), metaplastic carcinoma spindle cell (P < 0.01), metaplastic carcinoma squamous cell and secretory carcinoma.

    “In breast tissue, the pan-cancer model outperforms the Paige Breast model overall and especially on some rare histological variants, including adenoid cystic carcinoma, carcinoma with apocrine differentiation (P < 0.05), metaplastic carcinoma spindle cell (P < 0.01), metaplastic carcinoma squamous cell and the exceptionally unusual secretory carcinoma.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    From a total of 1,038 samples (1,087 WSIs), 26.2% have FGFR3 alterations.

    “From the total of 1,038 samples (1,087 WSIs), 26.2% have FGFR3 alterations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The Colon-MSI dataset comprises 2,698 CRC samples with 288 MSI-H/dMMR positive cases, uses both IHC and MSK-IMPACT sequencing for dMMR and MSI-H detection, and prioritizes IHC results when both test outcomes are available.

    “The Colon-MSI dataset, comprising 2,698 CRC samples with 288 MSI-H/dMMR positive cases, uses both IHC and MSK-IMPACT sequencing for dMMR and MSI-H detection, prioritizing IHC results when both test outcomes are available.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    The breast rare variants benchmark contains 23 cases of invasive ductal carcinoma or invasive lobular carcinoma as control, 75 cases of rare variants, and 392 benign cases (total 5,031 slides).

    “This dataset contains 23 cases of invasive ductal carcinoma or invasive lobular carcinoma (as control), 75 cases of rare variants”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Common cancer comprises 14,179 slides from 3,547 specimens, and rare cancer comprises 8,753 slides from 2,595 specimens.

    “Based on this definition, common cancer comprises 14,179 slides from 3,547 specimens originating in breast, prostate, lung, colon, skin, bladder, uterus, pancreas and H&N, and rare cancer comprises 8,753 slides from 2,595 specimens originating in liver, stomach, brain, ovary, cervix, testis and bone.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The breast product benchmark contains 190 slides with invasive cancer and 1,501 benign slides, labeled individually by a pathologist according to subtypes.

    “This dataset contains 190 slides with invasive cancer and 1,501 benign slides, labeled individually by a pathologist according to presence of atypical ductal hyperplasia, atypical lobular hyperplasia, lobular carcinoma in situ, ductal carcinoma in situ, invasive ductal carcinoma, invasive lobular carcinoma and/or other subtypes”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The CDH1 dataset comprises 1,077 ER+ primary breast cancer samples, of which 139 were positive and 918 were negative; 20 samples with monoallelic mutations were excluded.

    “The CDH1 dataset comprises a total of 1,077 estrogen receptor-positive (ER+) primary breast cancer samples, in which 139 were positive and 918 were negative. The remaining 20 samples with other types of variants—that is, monoallelic mutations—were excluded.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The testing dataset includes 15,622 slides from 3,033 specimens collected at MSKCC, in addition to 7,310 slides (3,109 specimens) sent to MSKCC from institutions around the world.

    “The testing dataset includes 15,622 slides from 3,033 specimens collected at MSKCC (denoted as ‘Internal’ in Fig. 2b), in addition to 7,310 slides (3109 specimens) sent to MSKCC from institutions around the world (‘External’ in Fig. 2b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Virchow marks a major increase in training data scale to 1.5 million WSIs, a volume over 3,000 times the size of ImageNet as measured by total number of pixels.

    “Virchow marks a major increase in training data scale to 1.5 million WSIs—a volume of data that is over 3,000 times the size of ImageNet31 as measured by the total number of pixels.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [31] Deng, J. et al. Imagenet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition 248–255 (IEEE, 2009). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    Pan-cancer detection generalizes across cancer types, including rare cancers, as well as on OOD data when using foundation model embeddings.

    “Overall, pan-cancer detection generalizes across cancer types, including rare cancers, as well as on OOD data when using foundation model embeddings.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Virchow embeddings generalize well to new or rare data and outperform the others consistently.

    “This demonstrates that Virchow embeddings generalize well to new or rare data and outperform the others consistently.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    Increased diversity helps to learn robust and rare features.

    “although we posit that increased diversity helps to learn robust and rare features. Indeed, it has been shown that training a model on multiple tissues or cancer variants can improve detection performance for each cancer53, as many morphological features are observed across cancers from different topographies54.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [53] Cheerla, A. & Gevaert, O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics 35, i446–i454 (2019). source

      Not checked in the free run.

    • cites [54] Noorbakhsh, J. et al. Deep learning-based cross-classifications reveal conserved spatial behaviors within tumor histological images. Nat. Commun. 11, 6367 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Study design and conductnone given

    Virchow embeddings were not fine-tuned, and models used simple aggregator architectures to make predictions.

    “Finally, we note that Virchow embeddings were not fine-tuned, and models used simple aggregator architectures to make predictions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductnone given

    The biomarker prediction results lacked sufficient statistical power to assess statistically significant differences across models; relative model performance was concluded from evaluating many different biomarker predictions.

    “We note that the biomarker prediction results lacked sufficient statistical power to assess statistically significant differences across models; instead, we conclude relative model performance from evaluating many different biomarker predictions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Recent pioneering works have utilized datasets ranging from 30,000 to 400,000 WSIs to train foundation models ranging in size from 28 million to 307 million parameters.

    “Despite the challenges in collecting large-scale datasets in the pathology domain, recent pioneering works have utilized datasets ranging from 30,000 to 400,000 WSIs to train foundation models ranging in size from 28 million to 307 million parameters35–42”
    Quote in the paperNot a substantial claim0 of 8 sources checked
    • cites [35] Wang, X. et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med. Image Anal. 81, 102559 (2022). source

      Not checked in the free run.

    • cites [36] Ciga, O., Xu, T. & Martel, A. L. Self supervised contrastive learning for digital histopathology. Mach. Learn. Appl. 7, 100198 (2022).

      Not checked in the free run.

    • cites [37] Filiot, A. et al. Scaling self-supervised learning for histopathology with masked image modeling. Preprint at https://doi.org/10.1101/2023.07.21.23292757 (2023). source

      Not checked in the free run.

    • cites [38] Azizi, S. et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nat. Biomed. Eng. 7, 1–24 (2023). source

      Not checked in the free run.

    • cites [39] Kang, M., Song, H., Park, S., Yoo, D. & Pereira, S. Benchmarking self-supervised learning on diverse pathology datasets. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 3344–3354 (IEEE, 2023). source

      Not checked in the free run.

    • cites [40] Dippel, J. et al. RudolfV: A foundation model by pathologists for pathologists. Preprint at https://doi.org/10.48550/arXiv.2401.04079 (2024). source

      Not checked in the free run.

    • cites [41] Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nat. Med. 30, 850–862 (2024). source

      Not checked in the free run.

    • cites [42] Campanella, G. et al. Computational pathology at health system scale–self-supervised foundation models from three billion images. Preprint at https://doi.org/10.48550/arXiv.2310.07033 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Virchow unlocks the ability to accurately and precisely detect unusual histological variants of cancer as well as biomarker status.

    “Indeed, Virchow unlocks the ability to accurately and precisely detect unusual histological variants of cancer as well as biomarker status, something that is difficult to achieve with cancer- or biomarker-specific training due to the limited amount of associated training data.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Trends in training data size may be oversimplified because they do not capture the tradeoff between increasing the number of WSIs versus tiles.

    “Trends in training data size may be oversimplified as they do not capture the tradeoff between increasing the number of WSIs versus tiles.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    These works demonstrate that image features produced with self-supervised learning of pathology images outperform image features trained on natural images and that performance improves with scale.

    “These works demonstrate that image features produced with self-supervised learning of pathology images outperform image features trained on natural images and that performance improves with scale.”
    Quote in the paperSubstantial claim0 of 8 sources checked
    • cites [35] Wang, X. et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med. Image Anal. 81, 102559 (2022). source

      Not checked in the free run.

    • cites [36] Ciga, O., Xu, T. & Martel, A. L. Self supervised contrastive learning for digital histopathology. Mach. Learn. Appl. 7, 100198 (2022).

      Not checked in the free run.

    • cites [37] Filiot, A. et al. Scaling self-supervised learning for histopathology with masked image modeling. Preprint at https://doi.org/10.1101/2023.07.21.23292757 (2023). source

      Not checked in the free run.

    • cites [38] Azizi, S. et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nat. Biomed. Eng. 7, 1–24 (2023). source

      Not checked in the free run.

    • cites [39] Kang, M., Song, H., Park, S., Yoo, D. & Pereira, S. Benchmarking self-supervised learning on diverse pathology datasets. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 3344–3354 (IEEE, 2023). source

      Not checked in the free run.

    • cites [40] Dippel, J. et al. RudolfV: A foundation model by pathologists for pathologists. Preprint at https://doi.org/10.48550/arXiv.2401.04079 (2024). source

      Not checked in the free run.

    • cites [41] Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nat. Med. 30, 850–862 (2024). source

      Not checked in the free run.

    • cites [42] Campanella, G. et al. Computational pathology at health system scale–self-supervised foundation models from three billion images. Preprint at https://doi.org/10.48550/arXiv.2310.07033 (2023). source

      Not checked in the free run.

  • May combine two claims. The quote only partly states this claim.

    Interpretationown data

    Virchow embeddings help predict biomarkers directly from H&E slides, reducing the need for targeted sequencing or IHC staining.

    “These findings underscore the potential of Virchow embeddings to accurately represent H&E histologic phenotypes, offering predictive insights into biomarkers that are traditionally identified through DNA extraction and MSK-IMPACT sequencing.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The prediction of biomarkers from standard H&E stained images can reduce the reliance on testing using additional methods and the associated substantial delays in returning results to patients.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DINO v.2 was shown to learn a similar semantic feature separation on natural images, allowing foreground/background separation and part annotation.

    “DINO v.2 was shown to learn a similar semantic feature separation on natural images, allowing foreground/background separation (for example, discriminating a bus or a bird from the background) as well as part annotation (for example, wheels versus windows in a bus)33.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Oquab, M. et al. DINOv2: Learning Robust Visual Features without Supervision. In Transactions on Machine Learning Research 2835–8856 (TMLR, 2024).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    In this work, algorithmic and training settings are largely based on what was successful in the natural image domain.

    “however, in this work, as in many other works in self-supervised learning for computational pathology, algorithmic and training settings are largely based on what was successful in the natural image domain.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Due to the rarity of these variants of cancers, rare variants prediction lacks the statistical power of the product datasets.

    “We note that due to the rarity of these variants of cancers, rare variants prediction lacks the statistical power of the product datasets.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    A rule-based natural language processing system was employed to establish the clinical cancer diagnosis at the specimen level.

    “To establish the clinical cancer diagnosis at the specimen level, a rule-based natural language processing system was employed.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    The Camelyon17-WILDS dataset is designed to test out-of-distribution generalization, with training data from three hospitals and validation and testing data from separate hospitals not represented in the training data.

    “Uniquely designed to test OOD generalization, the training set (335,996 images) is composed of data from three different hospitals, whereas the validation subset (34,904 images) and testing subset (85,054 images) each originate from separate hospitals not represented in the training data.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    If a validation set was not provided with the public dataset, 10% of the training data was randomly split out to make a validation set.

    “If one was not provided with the public dataset, we randomly split out 10% of the training data to make a validation set.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    AUC is a suitable general metric because it does not require selecting a threshold for the model's probability outputs.

    “AUC is a suitable general metric because it does not require selecting a threshold for the model’s probability outputs, something that may need tuning for different data subpopulations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    A validation set was used for all tasks.

    “A validation set was used for all tasks.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    Cancerous slides in the prostate rare variants benchmark are curated and labeled by a pathologist, and are appended with benign slides from synoptic reports collected at MSKCC.

    “Cancerous slides are curated and labeled by a pathologist, and are appended with slides from benign blocks determined from synoptic reports collected at MSKCC.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisereasoning

    Specificity at 95% sensitivity is informative because a clinical system must be not only sensitive but also specific in practice.

    “Specificity at 95% sensitivity is informative because a clinical system must be not only sensitive but also specific in practice.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    For testing set evaluation, the classifier checkpoint that achieved the lowest loss on the validation set was selected.

    “For testing set evaluation, the classifier checkpoint that achieved the lowest loss on the validation set was selected.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Due to the scale of training, the study has not been able to fully explore the effectiveness of data-balancing and -distillation strategies.

    “Due to the scale of training, our study has not been able to fully explore the effectiveness of data-balancing and -distillation strategies.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Such models must be extended to the slide level to achieve the data efficiency required for low-data tasks such as prediction of biomarkers, treatment response or clinical outcome.

    “Although scaling up the size of a tile-level foundation model may improve performance, it is likely that such models must be extended to the slide level to achieve the data efficiency required for low-data tasks such as the prediction of biomarkers, treatment response or clinical outcome.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductnone given

    The pan-cancer detection models are evaluated on a combination of data sourced from MSKCC and external institutions, and none of the patients in the evaluation set were seen during training.

    “The pan-cancer detection models are evaluated on a combination of data sourced from MSKCC and external institutions. None of the patients in the evaluation set were seen during training.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Screening for EGFR oncogenic mutations in non-small cell lung cancer is essential to determine eligibility for targeted therapies in late-stage non-small cell lung cancer.

    “The EGFR oncogenic mutation screening in non-small cell lung cancer is essential to determine eligibility for targeted therapies in late stage non-small cell lung cancer67.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [67] Kalemkerian, G. P. et al. Molecular testing guideline for the selection of patients with lung cancer for treatment with targeted tyrosine kinase inhibitors: American Society of Clinical Oncology endorsement of the College of American Pathologists/international Association for the Study of Lung Cancer/Association for Molecular Pathology Clinical Practice guideline update. J. Clin. Oncol. https://doi.org/10.1200/JCO.2017.76.7293 (2018). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data

    Increasing the number of model parameters remains a salient axis to explore.

    “Overall, our investigation into scaling behavior suggests that increasing the number of model parameters remains a salient axis to explore.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    Each slide in the BLN dataset has been labeled by a pathologist according to presence of invasive cancer, and the largest tumor on the slide is measured to categorize the tumor into macrometastasis, micrometastasis or infiltrating tumor cells.

    “Each slide has been labeled by a pathologist according to presence of invasive cancer, and the largest tumor on the slide is measured to categorize the tumor into macrometastasis, micrometastasis or infiltrating tumor cells.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Labels for the prostate blocks are extracted from synoptic reports collected at MSKCC.

    “Labels for the blocks are extracted from synoptic reports collected at MSKCC.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    A cut-off for FGA was established that enriched for TP53 mutations in the distribution of ovarian cancer cases.

    “A cut-off for FGA was established that enriched for TP53 mutations in the distribution of ovarian cancer cases.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Embeddings are generated at the tile level using ×20 magnification (0.5 mpp) as opposed to the slide level and therefore require training an aggregation model.

    “As with most histopathology self-supervised models, embeddings are generated at the tile level using ×20 magnification (0.5 mpp) as opposed to the slide level and therefore require training an aggregation model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    Due to the small biomarker dataset sizes, the learning rate was not chosen on a validation set to evaluate generalization; the reported best test AUC scores serve as a benchmark across tile embedding types and estimate the best possible biomarker performance for each type.

    “Due to the small biomarker dataset sizes, the learning rate was not chosen on a validation set to evaluate generalization; rather, this serves as a benchmark across the different types of tile embeddings (Virchow, UNI, Phikon and CTransPath), yielding an estimate of the best possible biomarker performance for each type.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Anecdotal experience from pathologists suggested there may be a morphological signal for FGFR alterations.

    “Anecdotal experience from pathologists suggested there may be a morphological signal for FGFR alterations66.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [66] Al-Ahmadie, H. A. et al. Somatic mutation of fibroblast growth factor receptor-3 (FGFR3) defines a distinct morphological subtype of high-grade urothelial carcinoma. J. Pathol. 224, 270–279 (2011). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The biallelic loss of the CDH1 gene is strongly correlated with lobular breast cancer and a distinct histologic phenotype and biologic behavior.

    “The biallelic loss of cadherin 1 (CDH1) gene (encoding E-cadherin) is strongly correlated with lobular breast cancer and a distinct histologic phenotype and biologic behavior65.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [65] Berx, G. & Van Roy, F. The E-cadherin/catenin complex: an important gatekeeper in breast cancer tumorigenesis and malignant progression. Breast Cancer Res. 3, 289–293 (2001). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Each biomarker dataset was randomly split into training and testing subsets, ensuring no patient overlap.

    “We randomly split each dataset into training and testing subsets, ensuring no patient overlap, as shown in Supplementary Table 3.1.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Each slide in the pan-tissue product benchmark is individually inspected by a pathologist and labeled according to presence of invasive cancer.

    “Each slide is individually inspected by a pathologist and labeled according to presence of invasive cancer.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Self-supervised learning is enabling in-domain transfer by removing the label requirement, driving a second wave of scaling to tens of thousands of WSIs.

    “Now self-supervised learning is enabling in-domain transfer by removing the label requirement, driving a second wave of scaling to tens of thousands of WSIs to inform image representation35–39,56.”
    Quote in the paperSubstantial claim0 of 6 sources checked
    • cites [35] Wang, X. et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med. Image Anal. 81, 102559 (2022). source

      Not checked in the free run.

    • cites [36] Ciga, O., Xu, T. & Martel, A. L. Self supervised contrastive learning for digital histopathology. Mach. Learn. Appl. 7, 100198 (2022).

      Not checked in the free run.

    • cites [37] Filiot, A. et al. Scaling self-supervised learning for histopathology with masked image modeling. Preprint at https://doi.org/10.1101/2023.07.21.23292757 (2023). source

      Not checked in the free run.

    • cites [38] Azizi, S. et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nat. Biomed. Eng. 7, 1–24 (2023). source

      Not checked in the free run.

    • cites [39] Kang, M., Song, H., Park, S., Yoo, D. & Pereira, S. Benchmarking self-supervised learning on diverse pathology datasets. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 3344–3354 (IEEE, 2023). source

      Not checked in the free run.

    • cites [56] Chen, R. J. et al. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 16123–16134 (IEEE, 2022). source

      Not checked in the free run.

66 of 66

  • The paper relies on this without support.

    The prediction of biomarkers from standard H&E stained images can reduce the reliance on testing using additional methods and the associated substantial delays in returning results to patients.

    “The prediction of biomarkers from standard H&E stained images can reduce the reliance on testing using additional methods and the associated substantial delays in returning results to patients (Fig. 4a).”

  • The paper relies on this without support.

    The success of such applications depends on the ability to model the diverse patterns observed in pathology images.

    “The success of such applications depends on the ability to model the diverse patterns observed in pathology images.”

  • The paper relies on this without support.

    The value of a pathology foundation model is twofold: generalizability and training data efficiency.

    “The value of a pathology foundation model is twofold: generalizability and training data efficiency.”

  • The paper relies on this without support.

    In clinical practice, most biopsy samples are benign.

    “In clinical practice, where most biopsy samples are benign, a pan-cancer detection system can prioritize cases to help reduce diagnostic turnaround.”

  • The paper relies on this without support.

    Histopathology data differs from natural image data in key ways: the long-tailed distribution of pathologic entities and histological structures, the lack of object scale diversity and the restricted color space.

    “Histopathology data differs from natural image data in key ways: the long-tailed distribution of pathologic entities and histological structures, the lack of object scale diversity and the restricted color space.”

  • The paper relies on this without support.

    Model performance critically depends on the model’s ability to capture image features.

    “This process, called transfer learning, was motivated by the observation that model performance critically depends on the model’s ability to capture image features.”

  • The paper relies on this without support.

    In-domain transfer learning was not possible given the limited availability of labeled pathology datasets.

    “In-domain transfer learning was not possible given the limited availability of labeled pathology datasets.”

  • The paper relies on this without support.

    AR amplification or overexpression is found in 30%–50% of castration-resistant prostate cancers and is associated with resistance to androgen deprivation therapy.

    “The AR amplification/overexpression was found in 30%–50% of castration resistant prostate cancers and was associated with resistance to androgen deprivation therapy.”

  • The paper relies on this without support.

    HER2 overexpression and/or amplification are more heterogeneous in gastric cancer compared to breast cancer.

    “Human epidermal growth factor receptor 2 (HER2) overexpression and/or amplification are much more heterogeneous in gastric cancer compared to breast cancer.”

  • The paper relies on this without support.

    Approximately 20% of gastric cancer patients have HER2 overexpression or high-level amplification.

    “Approximate 20% of gastric cancer patients are found to correlate with HER2 overexpression/high-level amplification, and they would be likely to benefit from treatment with an anti-HER2 antibody therapy.”

  • The paper relies on this without support.

    Gastric cancer patients with HER2 overexpression or high-level amplification would likely benefit from anti-HER2 antibody therapy.

    “Approximate 20% of gastric cancer patients are found to correlate with HER2 overexpression/high-level amplification, and they would be likely to benefit from treatment with an anti-HER2 antibody therapy.”

  • The paper relies on this without support.

    PTEN is the most frequently mutated tumor suppressor gene in endometrial cancer.

    “PTEN is the most frequently mutated tumor suppressor gene in endometrial cancer.”

  • The paper relies on this without support.

    The presence of PTEN mutation is significantly associated with poorer prognosis in survival and disease recurrence.

    “The presence of PTEN mutation showed to be significantly associated with poorer prognosis in survival and disease recurrence.”

  • The paper relies on this without support.

    RET mutations are highly associated with medullary thyroid cancer.

    “RET mutations were highly associated with medullary thyroid cancer, which accounts for about 5–10% of all thyroid cancer.”

  • The paper relies on this without support.

    Medullary thyroid cancer accounts for about 5–10% of all thyroid cancer.

    “RET mutations were highly associated with medullary thyroid cancer, which accounts for about 5–10% of all thyroid cancer.”

  • The paper relies on this without support.

    Screening RET oncogenic mutations plays an important role in diagnosis and prognosis of medullary thyroid cancer.

    “Screening RET oncogenic mutations plays an important role in diagnosis and prognosis of medullary thyroid cancer.”

  • The paper relies on this without support.

    BRAF is one of the most frequently mutated genes in melanoma.

    “BRAF is one of the most frequently mutated genes in melanoma, and V600E mutation is the most common variant, which leads to constitutive activation of the BRAF/MEK/ERK signaling pathway.”

  • The paper relies on this without support.

    V600E is the most common BRAF variant in melanoma.

    “BRAF is one of the most frequently mutated genes in melanoma, and V600E mutation is the most common variant, which leads to constitutive activation of the BRAF/MEK/ERK signaling pathway.”

  • The paper relies on this without support.

    BRAF V600E mutation leads to constitutive activation of the BRAF/MEK/ERK signaling pathway.

    “BRAF is one of the most frequently mutated genes in melanoma, and V600E mutation is the most common variant, which leads to constitutive activation of the BRAF/MEK/ERK signaling pathway.”

  • The paper relies on this without support.

    Targeted therapy with BRAF inhibitors shows better survival outcome in patients with BRAF V600-mutated melanoma.

    “Targeted therapy with BRAF inhibitors showed better survival outcome in patients with BRAF V600-mutated melanoma.”

  • The paper relies on this without support.

    High-grade serous ovarian cancer is characterized by high prevalence of TP53 mutations and genome instability with widespread genetic alteration.

    “High-grade serous ovarian cancer is characterized by high prevalence of TP53 mutations and genome instability with widespread genetic alteration.”
21 of 21

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Weakly-supervised deep learning models can be trained to perform various downstream computational pathology tasks such as tumor subtyping and prognosis without any region or pixel-level annotations.

    “To overcome these limitations, several research groups have alternatively proposed the use of weakly-supervised deep learning models, which can be trained to perform various downstream computational pathology tasks such as tumor subtyping and prognosis32–37 without any region or pixel-level annotations.”

    From Predicting the tumor microenvironment composition and immunotherapy response in non-small cell lung cancer from digital histopathology images (2024) · cites this paper as [36]

    The Discussion (page unavailable) says multiple-instance learning with labels for groups of slides enables clinically relevant diagnostics, supporting the claim that region- or pixel-level training annotations are unnecessary for some tasks. The Results (page unavailable) describe weakly supervised models for cancer detection and biomarker prediction, but the paper does not establish prognosis prediction as a demonstrated task; the Discussion treats clinical-outcome prediction as a future low-data application. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • UNI, Prov-GigaPath, and Virchow were pretrained solely on pathology image data to perform cancer classification, diagnosis, and prognosis.

    “UNI [191], Prov-GigaPath [192], and Virchow [193] were pretrained solely on pathology image data to perform cancer classification, diagnosis, and prognosis.”

    From Challenges in AI-driven Biomedical Multimodal Data Fusion and Analysis (2025) · cites this paper as [193]

    The Main section (page unavailable) says Virchow was trained on H&E pathology slides, and the Results section reports cancer detection using its embeddings. The Discussion describes clinical-outcome prediction as a potential future task, not a demonstrated prognosis application. This verdict concerns Virchow, the model attached to this citation, rather than the separately cited UNI and Prov-GigaPath. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • GMA is widely used to assess the performance of pathology foundation models.

    “In fact GMA is widely used to assess the performance of pathology foundation models23,25,26.”

    From A clinical benchmark of public self-supervised pathology foundation models (2025) · cites this paper as [25]

  • Foundation models have shown impressive performance across medical fields such as ophthalmology, radiology and pathology.

    “Built on rich knowledge representations, these models have shown impressive performance across medical fields such as ophthalmology20, radiology21 and pathology22–25.”

    From A Multimodal Vision Foundation Model for Clinical Dermatology (2025) · cites this paper as [24]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • CoNSeP feature analysis is assigned conflicting Fig. 5 panels

    Minor

    The text places the feature-separation examples in Fig. 5d, but the figure caption assigns embedding-feature visualization to panel c and describes no panel d.

  • Rare-variant results are assigned conflicting Fig. 3 panels

    Minor

    The text directs readers to Fig. 3d for rare-variant performance, but the figure caption assigns rare-variant performance to panel a and failure examples to panel d.

  • Pan-cancer comparison counts three models but names four

    Minor

    The statistical-analysis section says three models are compared, whereas the Results give pan-cancer AUCs for four models.

  • TCGA TIL split exceeds the stated dataset size

    Minor

    The stated training, validation and testing counts sum to 342,698 images, not 304,097.

    Reported 304097 · recomputed 342698

  • Biomarker AUC results are assigned to the positive-fraction panel

    Minor

    Methods says the test AUC scores are in Fig. 4b, while the figure caption identifies panel b as positive-case fractions and panel d as AUC performance.

5 of 5

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. In Proc. 37th International Conference on Machine Learning (eds Daumé, H. & Singh, A.) 1597–1607 (JMLR.org, 2020).
  2. Zhou, J. et al. Image BERT Pre-training with Online Tokenizer. In The Tenth International Conference on Learning Representations https://openreview.net/forum?id=ydopy-e6Dg (OpenReview.net, 2022).
  3. Dosovitskiy, A. et al. An image is worth 16 × 16 words: transformers for image recognition at scale. In The Ninth International Conference on Learning Representations https://openreview.net/forum?id=YicbFdNTTy (OpenReview.net, 2021).
  4. Ciga, O., Xu, T. & Martel, A. L. Self supervised contrastive learning for digital histopathology. Mach. Learn. Appl. 7, 100198 (2022).
  5. Casson, A. et al. Joint breast neoplasm detection and subtyping using multi-resolution network trained on large-scale H&E whole slide images with weak labels. In Proc. Medical Imaging with Deep Learning (eds Oguz, I. et al.) 18–38 (JMLR, 2024).
  6. Ilse, M., Tomczak, J. & Welling, M. Attention-based deep multiple instance learning. In Proc. 35th International Conference on Machine Learning (eds Dy, J. & Krause, A.) 2127–2136 (JMLR, 2018).
  7. Tarvainen, A. & Valpola, H. Mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results. In Proc. 31st Conference on Neural Information Processing Systems (eds von Luxburg, U. et al.) 1195–1204 (ACM, 2017).
  8. Loshchilov, I. & Hutter, F. Decoupled weight decay regularization. In n The Seventh International Conference on Learning Representations https://openreview.net/forum?id=Bkg6RiCqY7 (OpenReview.net, 2019).

Cited by

Sorted by citations

Related

Authors
Vorontsov, E., Bozkurt, A., Casson, A., Shaikovski, G., Zelechowski, M., Severson, K., Zimmermann, E., Hall, J., Tenenholtz, N., Fusi, N., Yang, E., Mathieu, P., van Eck, A., Lee, D., Viret, J., Robert, E., Wang, Y. K., Kunz, J. D., Lee, M. C. H., Bernhard, J. H., Godrich, R. A., Oakley, G., Millar, E., Hanna, M., Wen, H., Retamero, J. A., Moye, W. A., Yousfi, R., Kanan, C., Klimstra, D. S., Rothrock, B., Liu, S., & Fuchs, T. J.
Affiliations
  • Microsoft (United States) (US)
  • Memorial Sloan Kettering Cancer Center (US)
  • St George Hospital (AU)
  • University of Rochester (US)
Corresponding author
Liu, Siqi
ORCID
Venue
Nature Medicine
Publisher
Springer Science and Business Media LLC
Published
2024-07-22
Volume
30
Issue
10
Pages
2924-2935
ISSN
  • 1078-8956 (ISSN-L)
  • 1546-170X
DOI
10.1038/s41591-024-03141-0
PMC
PMC11485232
OpenAlex
W4400889241
Semantic Scholar
7c36a73d4bd4b8a96aa79b78d0c751858dde39b8
Cited by
572
References
67
Citations by year
2026: 332 · 2025: 212 · 2024: 28
FWCI
83.87
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
U.S. Department of Health & Human Services | NIH | National Cancer Institute (P30CA008748)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Radiomics and Machine Learning in Medical Imaging
Keywords
foundation models, computational pathology, histopathological images, rare tumor, Virchow, biomarker prediction, cell identification, clinical decision support, limited training data
MeSH terms
Artificial Intelligence, Humans, Neoplasms, Pathology, Clinical, ROC Curve, Biomarkers, Tumor, Computational Biology, Rare Diseases, Precision Medicine, Neoplasm Grading

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review