Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 59 of the 59 references it cites were found in a registry or an index; none of the 9 sources its claims cite is checked for support yet.

Benchmarking foundation models as feature extractors for weakly supervised computational pathology

Open access

Abstract

Numerous pathology foundation models have been developed to extract clinically relevant information. There is currently limited literature independently evaluating these foundation models on external cohorts and clinically relevant tasks to uncover adjustments for future improvements. Here we benchmark 19 histopathology foundation models on 13 patient cohorts with 6,818 patients and 9,528 slides from lung, colorectal, gastric and breast cancers. The models were evaluated on weakly supervised tasks related to biomarkers, morphological properties and prognostic outcomes. We show that a vision-language foundation model, CONCH, yielded the highest overall performance when compared with vision-only foundation models, with Virchow2 as close second, although its superior performance was less pronounced in low-data scenarios and low-prevalence tasks. The experiments reveal that foundation models trained on distinct cohorts learn complementary features to predict the same label, and can be fused to outperform the current state of the art. An ensemble combining CONCH and Virchow2 predictions outperformed individual models in 55% of tasks, leveraging their complementary strengths in classification scenarios. Moreover, our findings suggest that data diversity outweighs data volume for foundation models.

Overview

CoChat Summary

TL;DR

This study benchmarks 19 histopathology foundation models and 14 ensembles on 31 weakly supervised tasks across 13 cohorts with 6,818 patients and 9,528 slides from lung, colorectal, gastric and breast cancers. CONCH and Virchow2 show the best overall performance, and ensembling complementary models improves predictions, while data diversity appears more important than data volume.

Key findings

  • Across all 31 tasks, CONCH and Virchow2 had the highest mean AUROCs of 0.71, followed by Prov-GigaPath and DinoSSLPath at 0.69. By task type: morphology—CONCH 0.77, Virchow2 and DinoSSLPath 0.76; biomarkers—Virchow2 and CONCH 0.73, Prov-GigaPath 0.72; prognosis—CONCH 0.63, Virchow2 and BiomedCLIP 0.61.
  • The CONCH + Virchow2 ensemble produced a 1,792-dimensional vector with the highest AUROC of 71.9, versus 71.1 for CONCH, 70.9 for Virchow2 and 69.2 for Prov-GigaPath. Ensembling CONCH and Virchow2 outperformed individual models in 55% of tasks and reduced misclassifications versus CONCH by an average of 6.2% across five folds; the four top models ensemble achieved a mean AUROC 1.2% higher than CONCH.
  • Data diversity outweighed volume: correlations between downstream performance and pretraining dataset size/diversity ranged from r = 0.29–0.74, with significant correlations only for morphology with patient count (r = 0.73, P < 0.05) and tissue site diversity (r = 0.74, P < 0.05). CONCH outperformed BiomedCLIP despite far fewer image-caption pairs (1.1 million versus 15 million).
  • In low-data settings, Virchow2 led 8 tasks at n = 300 patients, PRISM led 9 at n = 150, and CONCH led 5 at n = 75. In low-prevalence biomarkers, Prov-GigaPath had the highest mean AUROC of 0.74, followed by Virchow (0.73) and CONCH (0.72). Slide encoders underperformed original tile embeddings in MIL, and ABMIL averaged 0.01 AUROC lower than the transformer-based aggregator.

Methods

  • Benchmark of 19 foundation models (12 vision-only, 3 vision-language, 4 slide encoders) and 14 ensembles on 31 weakly supervised tasks: 5 morphology, 19 biomarker and 7 prognostication tasks. Data included 13 cohorts, 6,818 patients and 9,528 slides from LUAD, LUSC, CRC, STAD and BRCA; TCGA was used for training, with external validation on CPTAC, DACHS, Kiel, Bern and IEO.
  • The STAMP pipeline v1.1.1 tessellated WSIs into 224 × 224 pixel tiles, extracted features with each foundation model, and trained a transformer-based aggregator using fivefold cross-validation, yielding 2,945 models (19 models × 31 tasks × 5 folds). ABMIL was evaluated as an alternative aggregator; metrics included AUROC, AUPRC, balanced accuracy, F1, DeLong’s test and Cohen’s kappa, with Grad-CAM heatmaps for interpretability.
  • Low-data experiments trained downstream models on randomly sampled cohorts of 300, 150 and 75 patients. Low-prevalence tasks included BRAF (10%), CIMP (13%) and MSI (14%) in CRC; EBV (8%) and M-status (7%) in STAD; and EGFR (11%) and STK11 (15%) in LUAD, evaluated in DACHS, Kiel and CPTAC-LUAD.

Limitations

  • Stated: evaluation covers only four cancer types; the study was limited to foundation models with research-accessible licences, excluding RudolphV and PLUTO; pen-mark artefacts were present in 90% of DACHS and 22% of Bern samples but had minimal impact because they were evenly distributed across classes; fine-tuning general-purpose models such as GPT-4o was outside scope; future work should expand tumour types, biomarkers and patient cohorts.
  • Stated/evident: correlation analyses made no adjustments for multiple comparisons, and most correlations between pretraining dataset characteristics and downstream performance were not statistically significant. Evident from the text: low-prevalence tasks were restricted to DACHS, Kiel and CPTAC-LUAD to avoid cancer type imbalance, so those findings may not generalize to other cohorts.

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

Integrity signals

Registry

Crossref · 10.1038/s41551-025-01516-3

Cites

59 of 59 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Experimental design of the study. Benchmarking of 19 histopathology foundation models using 13 cohorts and 31 tasks. a, Number of slides used from each of the 13 cohorts including 4 cancer types. b, About 9,528 haematoxylin and eosin (H&E) stained WSIs were preprocessed using the standardized STAMP19 pipeline. Feature extraction from the processed tiles was performed using 19 foundation models analysed in this study. The TCGA features were utilized for fivefold cross-validation with downstream transformer models on 31 classification tasks using STAMP. All models were subsequently applied to external features from CPTAC, Bern, Kiel, DACHS and IEO. The transformer architecture schematic shows layer normalization (Norm) and multi-headed self-attention (MSHA), followed by a MLP. c, All experiments were analysed using AUROCs, supplemented by AUPRC, Pearson’s correlation coefficient, DeLong’s test, balanced accuracy and F1 score. CONCH achieves the highest average AUROC across all tasks, followed by Virchow2, Prov-GigaPath and DinoSSLPath. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Attention heatmaps were generated for some slides to interpret differences between foundation models. Source data
Fig. 1 Experimental design of the study. Benchmarking of 19 histopathology foundation models using 13 cohorts and 31 tasks. a, Number of slides used from each of the 13 cohorts including 4 cancer types. b, About 9,528 haematoxylin and eosin (H&E) stained WSIs were preprocessed using the standardized STAMP19 pipeline. Feature extraction from the processed tiles was performed using 19 foundation models analysed in this study. The TCGA features were utilized for fivefold cross-validation with downstream transformer models on 31 classification tasks using STAMP. All models were subsequently applied to external features from CPTAC, Bern, Kiel, DACHS and IEO. The transformer architecture schematic shows layer normalization (Norm) and multi-headed self-attention (MSHA), followed by a MLP. c, All experiments were analysed using AUROCs, supplemented by AUPRC, Pearson’s correlation coefficient, DeLong’s test, balanced accuracy and F1 score. CONCH achieves the highest average AUROC across all tasks, followed by Virchow2, Prov-GigaPath and DinoSSLPath. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Attention heatmaps were generated for some slides to interpret differences between foundation models. Source dataCC BY · PubMed Central
Fig. 2 Performance of 19 pathology foundation models on 31 weakly supervised prediction tasks. a, AUROC scores of the four best foundation models, task-wise normalization. b, AUROC scores of the two best foundation models compared with the average prediction of the four best models (Avg-Pred) and the concatenated vectors of CONCH and Prov-GigaPath (Concat). c–e, Average AUROC scores of the five folds of each foundation model on morphology (c), biomarker (d) and prognosis (e) tasks. Task-wise normalization for better comparison of the foundation models. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. f, Stacked pie charts showing the number of tasks where each model achieved an average AUROC of >0.7, 0.6–0.7 or <0.6, grouped by task type. g, Average AUROC scores of the five folds using encoded tile embeddings from slide encoders versus the original tile embeddings. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source data
Fig. 2 Performance of 19 pathology foundation models on 31 weakly supervised prediction tasks. a, AUROC scores of the four best foundation models, task-wise normalization. b, AUROC scores of the two best foundation models compared with the average prediction of the four best models (Avg-Pred) and the concatenated vectors of CONCH and Prov-GigaPath (Concat). c–e, Average AUROC scores of the five folds of each foundation model on morphology (c), biomarker (d) and prognosis (e) tasks. Task-wise normalization for better comparison of the foundation models. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. f, Stacked pie charts showing the number of tasks where each model achieved an average AUROC of >0.7, 0.6–0.7 or <0.6, grouped by task type. g, Average AUROC scores of the five folds using encoded tile embeddings from slide encoders versus the original tile embeddings. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source dataCC BY · PubMed Central
Fig. 3 The impact of data diversity and volume on downstream weakly supervised classification performance. a–c, The impact of foundation model data diversity on downstream classification. Pearson’s correlation (two-sided) was used to assess associations between pretraining dataset characteristics and downstream performance. No adjustments were made for multiple comparisons. Correlation between the number of WSIs, patients and anatomic tissue sites in the pretraining dataset and the average AUROC for each downstream task type for all vision-only foundation models for which this data is available. Here, k denotes thousands. (a). Correlation between the number of ICPs (in millions, M) in the pretraining dataset and the average AUROC for each downstream task type for all vision-language foundation models (b). Performance of the respective cancer types correlated with the proportion of the cancer type in the pretraining dataset (c). All information that was available is shown (Supplementary Tables 6–8). d,e, Experiments with reduced downstream training sizes. Average AUROC scores across 29 tasks, trained with 75, 150 or 300 patients (d). Distribution of AUROC scores across all tasks for each model separately. Violin plots show kernel density estimates of AUROC scores, truncated at the observed range. The inner box marks the median and interquartile range (25th–75th percentiles), with whiskers extending to the most extreme values within 1.5× interquartile range (e). The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source data
Fig. 3 The impact of data diversity and volume on downstream weakly supervised classification performance. a–c, The impact of foundation model data diversity on downstream classification. Pearson’s correlation (two-sided) was used to assess associations between pretraining dataset characteristics and downstream performance. No adjustments were made for multiple comparisons. Correlation between the number of WSIs, patients and anatomic tissue sites in the pretraining dataset and the average AUROC for each downstream task type for all vision-only foundation models for which this data is available. Here, k denotes thousands. (a). Correlation between the number of ICPs (in millions, M) in the pretraining dataset and the average AUROC for each downstream task type for all vision-language foundation models (b). Performance of the respective cancer types correlated with the proportion of the cancer type in the pretraining dataset (c). All information that was available is shown (Supplementary Tables 6–8). d,e, Experiments with reduced downstream training sizes. Average AUROC scores across 29 tasks, trained with 75, 150 or 300 patients (d). Distribution of AUROC scores across all tasks for each model separately. Violin plots show kernel density estimates of AUROC scores, truncated at the observed range. The inner box marks the median and interquartile range (25th–75th percentiles), with whiskers extending to the most extreme values within 1.5× interquartile range (e). The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source dataCC BY · PubMed Central
Fig. 4 Divergence in tissue focus and predictive similarity among foundation models. a, Attention heatmap analysis for MSI-H classification in four different DACHS samples selected for correct predictions across selected foundation models. Thumbnails of the original WSIs and heatmaps of selected foundation models. b, Objective measure of similarity of prediction scores using Cohen’s kappa and majority vote across the five folds to binarize the predictions. Kappa scores of all combinations of foundation models tested in this study. c, Cohen’s kappa between the five folds of each foundation model. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source data
Fig. 4 Divergence in tissue focus and predictive similarity among foundation models. a, Attention heatmap analysis for MSI-H classification in four different DACHS samples selected for correct predictions across selected foundation models. Thumbnails of the original WSIs and heatmaps of selected foundation models. b, Objective measure of similarity of prediction scores using Cohen’s kappa and majority vote across the five folds to binarize the predictions. Kappa scores of all combinations of foundation models tested in this study. c, Cohen’s kappa between the five folds of each foundation model. The star indicates that Panakeia was tested on all tasks despite being specifically designed for BRCA and CRC. Source dataCC BY · PubMed Central
Extended Data Fig. 1 AUROCs, AUPRCs, balanced accuracy and F1-scores for all main experiments. A-D, Average AUROC (A), AUPRC (B), balanced accuracy (C) and F1 (D) scores of the five-folds of each foundation model on Morphology, Biomarker and Prognosis tasks. Source data
Extended Data Fig. 1 AUROCs, AUPRCs, balanced accuracy and F1-scores for all main experiments. A-D, Average AUROC (A), AUPRC (B), balanced accuracy (C) and F1 (D) scores of the five-folds of each foundation model on Morphology, Biomarker and Prognosis tasks. Source dataCC BY · PubMed Central
Extended Data Fig. 2 Average AUROCs sorted by cancer type and on scarce data tasks. Average AUROC scores of the five folds of each foundation model. Taskwise normalization for better comparison of the foundation models. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. A, The 31 tasks were grouped by cancer type (5 tasks for NSCLC, 5 tasks for BRCA, 8 tasks for STAD, 13 tasks for CRC). Models are sorted by average performance. B, Only tasks with rare positive cases (>15%) in the TCGA training cohort are shown. To avoid cancer type imbalance, these tasks are only evaluated in DACHS, Kiel and CPTAC LUAD. Source data
Extended Data Fig. 2 Average AUROCs sorted by cancer type and on scarce data tasks. Average AUROC scores of the five folds of each foundation model. Taskwise normalization for better comparison of the foundation models. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. A, The 31 tasks were grouped by cancer type (5 tasks for NSCLC, 5 tasks for BRCA, 8 tasks for STAD, 13 tasks for CRC). Models are sorted by average performance. B, Only tasks with rare positive cases (>15%) in the TCGA training cohort are shown. To avoid cancer type imbalance, these tasks are only evaluated in DACHS, Kiel and CPTAC LUAD. Source dataCC BY · PubMed Central
Extended Data Fig. 3 Comparison of STAMP and ABMIL. A, Difference in average AUROC scores between STAMP transformer-based aggregation and ABMIL across all tasks, calculated as the average over five cross-validation folds for each foundation model. Positive values indicate superior performance of STAMP. B, Difference in average macro-AUC scores between STAMP and ABMIL for selected foundation models under reduced downstream training dataset conditions, as shown in Extended Data Fig. 5. This compares the relative performance of both methods in low-data scenarios. Source data
Extended Data Fig. 3 Comparison of STAMP and ABMIL. A, Difference in average AUROC scores between STAMP transformer-based aggregation and ABMIL across all tasks, calculated as the average over five cross-validation folds for each foundation model. Positive values indicate superior performance of STAMP. B, Difference in average macro-AUC scores between STAMP and ABMIL for selected foundation models under reduced downstream training dataset conditions, as shown in Extended Data Fig. 5. This compares the relative performance of both methods in low-data scenarios. Source dataCC BY · PubMed Central
Extended Data Fig. 4 Performance Comparison of Model Ensembles and Single-Model Baselines Using DeLong’s Test. A, AUROC scores for each model and ensemble approach are shown, averaging predictions across five folds for individual models and five or ten folds for ensembles. Two ensembling approaches were used: taking the average prediction scores of downstream models trained on different foundation model backbones (prefix Avg) and concatenating feature vectors from different backbones to create a single downstream model (prefix Concat). The “Lauren” task was excluded as it’s not a binary classification. B-C, P-values from two-sided DeLong’s tests comparing CONCH (B) or Virchow2 (C) with other models and ensembles. No correction for multiple testing was applied; alpha was set to 0.05. Source data
Extended Data Fig. 4 Performance Comparison of Model Ensembles and Single-Model Baselines Using DeLong’s Test. A, AUROC scores for each model and ensemble approach are shown, averaging predictions across five folds for individual models and five or ten folds for ensembles. Two ensembling approaches were used: taking the average prediction scores of downstream models trained on different foundation model backbones (prefix Avg) and concatenating feature vectors from different backbones to create a single downstream model (prefix Concat). The “Lauren” task was excluded as it’s not a binary classification. B-C, P-values from two-sided DeLong’s tests comparing CONCH (B) or Virchow2 (C) with other models and ensembles. No correction for multiple testing was applied; alpha was set to 0.05. Source dataCC BY · PubMed Central
Extended Data Fig. 5 Model performance with reduced downstream training dataset. Mean AUROC across all five folds on 29 tasks for all foundation models trained with a reduced downstream dataset of 75 (A), 150 (B), or 300 patients (C). Patients were randomly selected from the TCGA cohorts, ensuring the ground truth was defined for all analyzed tasks. The tasks Lauren in Kiel and Bern were excluded due to insufficient patient numbers. Source data
Extended Data Fig. 5 Model performance with reduced downstream training dataset. Mean AUROC across all five folds on 29 tasks for all foundation models trained with a reduced downstream dataset of 75 (A), 150 (B), or 300 patients (C). Patients were randomly selected from the TCGA cohorts, ensuring the ground truth was defined for all analyzed tasks. The tasks Lauren in Kiel and Bern were excluded due to insufficient patient numbers. Source dataCC BY · PubMed Central
Extended Data Fig. 6 High-performance vs. low-performance tasks. A, Average AUROC scores across 15 high-performance and 16 low-performance tasks. Tasks were selected by including only those where at least one foundation model achieved an average AUROC over 0.75 and all others in low-performance tasks. B-C, The performance of each foundation model is listed. The final row presents the overall average AUROC for each model. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. Source data
Extended Data Fig. 6 High-performance vs. low-performance tasks. A, Average AUROC scores across 15 high-performance and 16 low-performance tasks. Tasks were selected by including only those where at least one foundation model achieved an average AUROC over 0.75 and all others in low-performance tasks. B-C, The performance of each foundation model is listed. The final row presents the overall average AUROC for each model. Tasks are sorted by their mean AUROC across all models, while models are sorted by their mean AUROC across all tasks. Source dataCC BY · PubMed Central
Extended Data Fig. 7 Attention heatmaps of slides with large variations in prediction scores. A-C, Attention Heatmap Analysis for Kiel EBV status (A), DACHS CRC sidedness (B) and CPTAC-BRCA ESR1 expression (C). Classification in 12 different WSIs selected for diverse prediction scores across the foundation models. Thumbnails of the original whole slide images (WSIs) and heatmaps of all foundation models are shown.
Extended Data Fig. 7 Attention heatmaps of slides with large variations in prediction scores. A-C, Attention Heatmap Analysis for Kiel EBV status (A), DACHS CRC sidedness (B) and CPTAC-BRCA ESR1 expression (C). Classification in 12 different WSIs selected for diverse prediction scores across the foundation models. Thumbnails of the original whole slide images (WSIs) and heatmaps of all foundation models are shown.CC BY · PubMed Central
Extended Data Fig. 8 Attention heatmaps of slides that all models predicted well. A-C, Attention Heatmap Analysis for Kiel N status (A), NSCLC subtyping (B) and CPTAC-BRCA ESR1 expression (C). Classification in four different samples per cohort selected for correct predictions across almost all foundation models. Thumbnails of the original whole slide images (WSIs) and heatmaps of all foundation models are shown.
Extended Data Fig. 8 Attention heatmaps of slides that all models predicted well. A-C, Attention Heatmap Analysis for Kiel N status (A), NSCLC subtyping (B) and CPTAC-BRCA ESR1 expression (C). Classification in four different samples per cohort selected for correct predictions across almost all foundation models. Thumbnails of the original whole slide images (WSIs) and heatmaps of all foundation models are shown.CC BY · PubMed Central
Extended Data Fig. 9 AUROC scores across all foundation models and ensembles. AUROC scores for all foundation models, foundation model variations, and multiple ensemble approaches. Prov-GigaPath-T are the regular tile embeddings, Prov-GigaPath-S are the tile embeddings encoded by the GigaPath slide encoder. Virchow(-2)-CLS contained only class tokens, with Virchow(-2)-CLS + MPT representing the version with class and mean patch tokens combined. Multiple experiments were conducted using concatenated feature vectors combining features from CONCH, Virchow2, Prov-GigaPath, DinoSSLPath, H-optimus-0 and UNI. For the same combinations, average prediction scores were calculated. These scores were used to evaluate the performance of combined predictions. Source data
Extended Data Fig. 9 AUROC scores across all foundation models and ensembles. AUROC scores for all foundation models, foundation model variations, and multiple ensemble approaches. Prov-GigaPath-T are the regular tile embeddings, Prov-GigaPath-S are the tile embeddings encoded by the GigaPath slide encoder. Virchow(-2)-CLS contained only class tokens, with Virchow(-2)-CLS + MPT representing the version with class and mean patch tokens combined. Multiple experiments were conducted using concatenated feature vectors combining features from CONCH, Virchow2, Prov-GigaPath, DinoSSLPath, H-optimus-0 and UNI. For the same combinations, average prediction scores were calculated. These scores were used to evaluate the performance of combined predictions. Source dataCC BY · PubMed Central
Extended Data Fig. 10 Cohen’s kappa scores across all ensembles and their individual model components. Objective measure of similarity of prediction scores using Cohen’s Kappa and majority vote across the five folds to binarize the predictions. The concatenated versions of CONCH, Virchow2 (V2), Prov-GigaPath (GP), H-optimus-0 (HO0), UNI and DinoSSLPath (Dino) and their single model counterparts are shown. Source data
Extended Data Fig. 10 Cohen’s kappa scores across all ensembles and their individual model components. Objective measure of similarity of prediction scores using Cohen’s Kappa and majority vote across the five folds to binarize the predictions. The concatenated versions of CONCH, Virchow2 (V2), Prov-GigaPath (GP), H-optimus-0 (HO0), UNI and DinoSSLPath (Dino) and their single model counterparts are shown. Source dataCC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 2 Not valid · 7 Not checked · 83 Own results and interpretations

44 more claims

Create a free account to see all 92 claims and to verify the 9 references still open.

See all claims — free account
  • It may state two separate claims as one. 25
  • Probably not a substantial claim. 21
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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?

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.

1 of 1

Citations

Cites

Show all 56 references

Research works printed but not linked: 3

  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. III & Singh, A.) Vol. 119, 1597–1607 (PMLR, 2020).
  2. Dosovitskiy, A. et al. An image is worth 16×16 words: transformers for image recognition at scale. In Proc. International Conference on Learning Representations (ICLR, 2021).
  3. Ilse, M., Tomczak, J. M. & Welling, M. Attention-based deep multiple instance learning. In Proc. 35th International Conference on Machine Learning Vol. 80, 2127–2136 (PMLR, 2018)

Cited by

See all 81

Metadata

Authors
Neidlinger, P., El Nahhas, O. S. M., Muti, H. S., Lenz, T., Hoffmeister, M., Brenner, H., van Treeck, M., Langer, R., Dislich, B., Behrens, H. M., Röcken, C., Foersch, S., Truhn, D., Marra, A., Saldanha, O. L., & Kather, J. N.
Affiliations
  • University Hospital Carl Gustav Carus (DE)
  • Helmholtz-Zentrum Dresden-Rossendorf (DE)
  • German Cancer Research Center (DE)
  • Johannes Kepler University of Linz (AT)
  • University of Bern (CH)
  • University Hospital Schleswig-Holstein (DE)
  • University Medical Center of the Johannes Gutenberg University Mainz (DE)
  • Universitätsklinikum Aachen (DE)
  • European Institute of Oncology (IT)
  • Heidelberg University (DE)
Corresponding author
Kather, Jakob Nikolas
Venue
Nature Biomedical Engineering
Published
2025-10-01
Volume
10
Issue
6
Pages
1113-1123
ISSN
2157-846X (ISSN-L)
OpenAlex
W4414705991
Cited by
80
References
56
Citations by year
2026: 69 · 2025: 11
FWCI
31.97
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Biomedical Engineering (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • Archivio Istituzionale della Ricerca (Universita Degli Studi Di Milano) (submittedVersion, other-oa)
  • Multimedialen Archiv und Publikationsserver der Christian-Albrechts-Universität zu Kiel (Christian-Albrechts-Universität zu Kiel) (submittedVersion, cc-by)
Funders
National Cancer Institute (R01CA263318)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Radiomics and Machine Learning in Medical Imaging
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
feature extraction, weakly supervised learning, computational pathology, vision-language models, conch, model benchmarking, ensemble learning, biomarker prediction, prognostic outcome, lung cancer, colorectal cancer, gastric cancer, breast cancer, data diversity
MeSH terms
Algorithms, Female, Humans, Neoplasms, Prognosis, Computational Biology, Benchmarking

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.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Benchmarking foundation models as feature extractors for weakly supervised computational pathology. https://app.cochat.ai/papers/10.1038/s41551-025-01516-3
The paper
Neidlinger, P., El Nahhas, O. S. M., Muti, H. S., Lenz, T., Hoffmeister, M., Brenner, H., van Treeck, M., Langer, R., Dislich, B., Behrens, H. M., Röcken, C., Foersch, S., Truhn, D., Marra, A., Saldanha, O. L., & Kather, J. N. (2025). Benchmarking foundation models as feature extractors for weakly supervised computational pathology. Nature Biomedical Engineering. https://doi.org/10.1038/s41551-025-01516-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.