A benchmark study of vision and pathology foundation models for computational pathology Authors: Rohan Bareja, Francisco Carrillo-Perez, Yuanning Zheng, Marija Pizurica, Tarak Nath Nandi, Lu Tian, Jeanne Shen, Ravi Madduri, Olivier Gevaert Published: 2026 ABSTRACT AbstractTo advance precision medicine in pathology, artificial intelligence (AI)-driven foundation models must generalize across diverse datasets, tissues, and clinical tasks. However, their comparative performance and generalizability in computational pathology remain incompletely characterized. Here, we benchmark 32 AI foundation models across four categories, including general vision models (VM), general vision-language models (VLM), pathology-specific vision models (Path-VM), and pathology-specific vision-language models (Path-VLM), using slide- and patch-level tasks from The Cancer Genome Atlas (TCGA), Clinical Proteomic Tumor Analysis Consortium (CPTAC), external benchmarking datasets, and out-of-domain datasets. Across TCGA tasks, Path-VMs consistently rank among the strongest performers. Evaluation across CPTAC and out-of-domain datasets reveals more nuanced generalization behavior, with model rankings showing modest but consistent shifts across datasets and task categories. Pairwise statistical comparisons indicate that differences among top-performing models are often small and task dependent. Path-VMs outperform Path-VLMs and remain competitive with VMs. Model size and pretraining dataset scale do not consistently predict downstream performance. Finally, late decision-level ensembling improves aggregate performance across external datasets and tissue types, highlighting complementary strengths across foundation models. PathBench: https://pathbench.stanford.edu/Subject terms: Computational models, Biomarkers, Machine learning, Cancer imaging, Pathology Introduction Pathology plays a crucial role in cancer research, providing insights that drive the understanding, diagnosis, and treatment of cancer1. It involves thorough examination of tissue specimens for tumor diagnosis and identification of important prognostic features such as tumor stage and grade, as well as biomarkers suggestive of treatment response2. Pathologists study digital pathology slides to analyze tumor histologic features, including architectural patterns and cellular morphology, in order to detect and classify tumors3. Consequently, oncologic pathology plays an integral role in personalized medicine and treatment selection. Computational pathology (CPath) has enhanced the field of pathology by integrating automated workflows and processes and advanced computational techniques1. It has the potential to reduce the workload of pathologists considerably and can aid in speeding up the overall diagnostic process. With the advent of machine learning and artificial intelligence, CPath has proven potential to facilitate detection, classification, and predictive modeling of cancer4. Furthermore, CPath has proven effective in assisting pathologists with tumor grading, staging, and subtyping5, all of which support diagnostic decision-making. It has also been employed as a computational technique to study patterns, segment tumors6, and extract intricate features, thereby enhancing cancer detection7. The use of Artificial Intelligence (AI) has been at the forefront of medical research for the past decade8. AI algorithms based on convolutional neural networks and vision transformers have significantly advanced computational pathology image analysis, especially in automated tumor detection and classification9. AI-driven predictive models can analyze pathology images along with clinical and genomic data to predict patient outcomes, including disease progression, recurrence risk, and treatment response10–12. Integrating gene expression, spatial omics, single-cell omics, and digital pathology data with AI enables a comprehensive multi-modal approach to cancer diagnosis and biomarker discovery, leveraging deep learning to uncover intricate spatial and molecular relationships within the tumor microenvironment13,14. AI models can automatically segment and delineate tumor regions, extract complex features, and perform risk based stratification15. Additionally, AI has accelerated Computational Pathology (CPath) by automating image analysis, streamlining workflows, integrating multimodal data16,17, and enhancing diagnostic accuracy18. Despite these advancements, several challenges persist, particularly regarding data limitations and model generalizability. Pathology datasets suffer from a shortage of labeled data, as manual pathologist annotations are time-consuming and cost-ineffective19. In contrast, large-scale datasets like ImageNet20 benefit from extensive labeled data, enabling supervised learning approaches to achieve high accuracy across diverse natural image tasks. To address this, Self-Supervised Learning (SSL) has emerged as a machine learning paradigm where models are trained on large-scale datasets to learn meaningful representations without the need for labeled examples21. This approach reduces the burden associated with data labeling and speeds up development of AI models. SSL based foundation models can be pre-trained on diverse and large unlabeled datasets22,23 to learn generalized patterns, then fine-tuned on specific pathology datasets of interest24. SSL enables these foundation models to adapt to different types of pathologies, domains, and tasks, improving their generalizability25,26. Consequently, there has been a large effort in developing general purpose pathology foundation models trained on millions of whole slide images26–28, which can be adapted to tumor classification, segmentation, and grading tasks with smaller amounts of labeled data. One of the first prominent models to be released was CTransPath29, which was pretrained on the TCGA30 and PAIP datasets31,32 and trained on approximately 15 million images. Lunit’s DINO-based ViT-Small model24 also utilized TCGA, along with internal datasets comprising over 32.6 million patches and 36,000 slides. Phikon33, another SSL model based on the iBOT framework, was pretrained on 40 million images from TCGA, covering 16 cancer types. More recently, several models such as UNI25, Virchow26, and Hibou27 have adopted the DINOv2 SSL method with ViT-Large architectures for their pretraining, using proprietary datasets. Next, pathology language models such as PLIP34, Quilt-Net35, and CONCH36, which are trained on both pathology slides and paired text, have demonstrated significant performance in vision-based tasks as well. Vision encoders like CLIP37, iBOT23 that are built into these models have shown competitive results in natural image classification and image retrieval tasks. Additionally, ImageNet pre-trained models have been widely used as feature extractors for digital pathology38–40. Despite these advancements in SSL, current foundation models may still be biased toward the datasets they are trained on, limiting their transferability across diverse institutions and staining protocols41. Additionally, biases can arise from imbalanced representation of tissue types and the predominance of certain cancer subtypes in training data, all of which can impact the model’s generalization and clinical applicability42,43. Many pathology foundation models rely on self-supervised pre-training using datasets like TCGA24,33,44 or large proprietary datasets from a single institution25,26, which can inadvertently result in domain-specific overfitting rather than true generalization across diverse clinical settings43. This is further compounded by the limited availability of large-scale, diverse pathology datasets that span multiple institutions, staining variations, and scanner types45. Unlike general computer vision SSL models that benefit from extensive public datasets such as ImageNet20 or LAION-5B46, pathology-specific SSL models often lack sufficient pre-training diversity, making them susceptible to dataset-specific artifacts. A systematic comparison is needed to evaluate pathology foundation models against both domain-specific SSL models and general computer vision SSL models, assessing their robustness across independent external datasets. Additionally, benchmarking should consider model generalization across diverse tissue types and cancer subtypes, ensuring that performance gains are not restricted to specific datasets. In this work, we compare general vision models (VM), general vision-language models (VLM), pathology-specific vision models (Path-VM), and pathology-specific vision-language models (Path-VLM), analyzing their generalization capability and task-specific strengths. Altogether, over 20 pathology or pathology-language models have been released, and several studies47 have used off-the-shelf ImageNet-trained models such as ResNet, DINO, and DINOv238,39,48,49 for medical imaging, including digital pathology. However, a comprehensive comparison of these models across diverse datasets and tasks remains lacking. While prior benchmarking studies50,51 have primarily compared pathology-specific encoders, our work expands the scope substantially by evaluating 32 models across four model categories: general vision (VM), general vision-language (VLM), pathology-specific vision models (Path-VM), and pathology-specific vision-language models (Path-VLM). We evaluate these models on 41 downstream tasks spanning 53 datasets and more than 17,500 whole-slide images. Our evaluation encompasses a wide range of diagnostic, molecular, transcriptomic, and pathway-level prediction tasks, including both organ-specific and pan-cancer formulations, and covers tissues such as brain, bladder, prostate, kidney, breast, and lung. In contrast to prior benchmarks51 that primarily focus on a limited set of cancer types or weakly supervised classification tasks, our work expands the scope of evaluation to include pathway activation, transcriptomic subtyping, and multi-organ generalization. We further assess model robustness by explicitly separating TCGA (in-domain) from non-TCGA evaluations and analyzing both slide-level and tile-level settings under a unified protocol; this benchmark provides a complementary perspective on tissue-specific performance and cross-domain generalization of pathology foundation models. All models are evaluated within a standardized experimental framework using frozen feature extraction with linear probing, enabling a controlled comparison of representation transfer. Finally, we include recently released pathology foundation models, including UNI2, H-optimus-mini52, and EXAONEPath53, and provide analyses of scaling behavior, tissue-specific specialization, and ensemble-based strategies for improving robustness across datasets and task types. This work aims to provide insights into the factors driving effective pathology foundation models, offering a foundation for optimizing their application in precision medicine. Results Benchmarking pathology-specific vision models: demonstrating competitive performance and tumor-specific strengths We conducted an extensive benchmarking study of 32 models across diverse domains, including general vision, general vision-language, pathology-specific vision, and pathology specific vision-language, evaluating them over 41 distinct tasks. To ensure consistent and concise terminology throughout the study, we assigned model abbreviations based on their training domains and modalities (Table 1): pathology-specific vision models (Path-VM; Table 2), and pathology-specific vision-language models (Path-VLM; Table 3), general vision models (VM; Table 4), general vision-language models (VLM; Table 5). Among these, 19 tasks were curated from the TCGA dataset, while 22 tasks were curated from non-TCGA datasets and included multiple tumor types such as brain, breast, bladder, and others. To better assess model generalization across datasets with different acquisition sources and potential pretraining overlap, we evaluated performance separately on TCGA and non-TCGA tasks, where the latter included CPTAC cohorts, external benchmarking datasets, and institution-specific out-of-domain datasets. We evaluated model performance using AUROC, AUPRC, Sensitivity, and Specificity as the performance metrics for a comprehensive assessment. Initially, we focused on 16 pathology-specific vision models (Path-VM, Table 2), tested across 41 tasks related to histologic subtyping, biomarker status prediction, molecular subtyping, and grading (Supplementary Tables 1 and 2). Table 1Abbreviations of model categoriesModel categoryModel category abbreviationTraining dataGeneral vision modelsVMNatural imagesGeneral vision language modelsVLMNatural images and textPathology-specific vision modelsPath-VMPathology-specific imagesPathology-specific vision language modelsPath-VLMPathology-specific images and text Abbreviations of model categories Table 2Pathology-specific vision model (Path-VM) summaryModel NameSSL methodModel ArchitectureParameters in millionNo. of WSI’s in millionNo. of patches in millionNo. of cancer/tissue typesData SourcecTranspathContrastive learningCNN + SwinT~27.0321532 cancer typesTCGA, PAIPKaikoDINOViT-base~85.029NA32 cancer typesTCGAHIPTDINOViT-small~21.0110433 cancer typesTCGAVirchowDINOv2Vit-huge~6321.52000NAProprietaryLunitDINOViT-small~21.03632.5NATCGA, TULIPUNIDINOv2ViT-large~3070.110020 tissue typesProprietary, GTExHibouDINOv2ViT-base~8511200NAProprietaryPhikoniBOTViT-base~850.0064016 cancer typesTCGAGPFMDINOv2ViT-large~3070.08619034 tissue typesTCGA, PAIP and 49 public datasetsH-optimus-ODINOv2Vit-giant~11000.5NANAProprietaryUNI2DINOv2Vit-huge~6320.35200NAProprietaryPhikon-v2DINOv2ViT-large~3070.05846030 cancer sitesTCGA, GTEx, CPTAC, TCIA datasetsVirchow2DINOv2Vit-huge~6323.1NA17 tissue typesProprietaryProv-GigaPathDINOv2Vit-giant~11000.171130031 tissue typesProprietaryH-optimus-miniDINOv2 distilledVit-base~85.0064316 tissue typesTCGAEXAONEPathDINOv1Vit-Base~85.03285-Public Pathology-specific vision model (Path-VM) summary Table 3Pathology-specific vision-language model (Path-VLM) summaryModel namePretraining methodModel architectureParameters in millionData sourcePLIPCLIPViT-base~85208k pathology image-text pairsQuiltNet-B16CLIPViT-base~851 M pathology image-text pairsBiomedCLIPCLIPViT-base~8514 M pathology image-text pairsMizero pubmedbertContrastive learningViT-small~2133k pathology image-text pairsMizero clinmedbertContrastive learningViT-small~2133k pathology image-text pairsCONCHCoCa92ViT-base~851.17 pathology image-text pairsTITANiBOT + CoCa92ViT-base~85336k whole slide images & 423k image captions Pathology-specific vision-language model (Path-VLM) summary Table 4General vision model (VM) summaryModel nameSSL methodModel architectureParameters in millionData sourceDino-S16DINOViT-small~21ImageNet (14 M)Dino-B16DINOViT-base~85ImageNet (14 M)Dinov2DINOv2ViT-large~300LVD-142MiBOT-B16iBOTViT-base~85ImageNet (14 M)iBOT-L16iBOTViT-large~307ImageNet (14 M) General vision model (VM) summary Table 5General vision-language model (VLM) summaryModel NamePretraining methodModel ArchitectureParameters in millionData sourceCLIP-B16CLIP (Contrastive learning)ViT-base~115400 million (image, text) pairsBLIP-B16-14MBLIPViT-base~85ImageNet (14 M)Align-baseEfficientNet + BERT (Contrastive learning)ViT-base~746700 million (image, text) pairsbeit3-L16iBOTViT-large~307ImageNet (14 M) General vision-language model (VLM) summary The performance evaluation revealed distinct insights through multiple approaches focused on AUROC. A mean performance of AUROC across all tasks (Fig. 1b) highlighted the competitive landscape among models, with Virchow2 (Zimmermann et al. 2024) achieving a mean AUROC of 0.830 (95% CI: 0.784–0.875), Prov-GigaPath (0.829, 95% CI: 0.784–0.875), H-optimus-0 (0.824, 95% CI: 0.776–0.873), UNI25 (0.824, 95% CI: 0.777–0.871), and UNI2 (0.824, 95% CI: 0.779–0.868), comprising the top five performers. Importantly, a consistent ranking pattern was also observed across non-TCGA tasks (22 tasks spanning CPTAC, out-of-domain slide-level cohorts, and external tile-level benchmarks; Fig. 1c). In this setting, Virchow2 (0.836, 95% CI: 0.767–0.906), UNI (0.835, 95% CI: 0.770–0.899), H-optimus-0 (0.832, 95% CI: 0.764–0.899), Prov-GigaPath (0.827, 95% CI: 0.757–0.897), and UNI2 (0.824, 95% CI: 0.751–0.897) again constituted the top-performing group. Task-specific rankings, assessed by identifying the top model per task based on AUROC, showed Virchow2 leading on 5 out of 19 TCGA-based tasks, followed by Prov-GigaPath (Xu et al. 2024) and H-optimus-0 each ranking first on 2 tasks, and UNI2 on 3 tasks (Fig. 1d). A similar pattern was observed across non-TCGA tasks (Fig. 1e, 22 tasks spanning CPTAC, out-of-domain slide-level cohorts, and external benchmarking datasets), where Virchow2 ranked first on 6 tasks, H-optimus-0 on 5 tasks, Prov-GigaPath on 2 tasks, and UNI and UNI2 each leading on 1 task. Benchmarking workflow and Path-VM performance across TCGA and non-TCGA tasks. a Benchmarking workflow. Schematic overview of data preprocessing, model feature extraction, evaluation across histopathology datasets, and performance comparison for tumor classification, molecular subtyping, and grading tasks. b Model performance on TCGA tasks. Path-VMs are sorted by mean AUROC across TCGA tasks. Colored circles represent individual task-level AUROC estimates, with each point corresponding to one model evaluated on one task. Black diamonds indicate mean AUROC, and horizontal error bars indicate 95% confidence intervals around the mean, calculated using a t-based standard error across task-level AUROC estimates. For AUROC, n = 19 task-level estimates per model across 16 Path-VMs. c Model performance on non-TCGA tasks. Path-VMs are sorted by mean AUROC across non-TCGA tasks. Points, means, and error bars are defined as in panel b. For AUROC, n = 22 task-level estimates per model across 16 Path-VMs. d Frequency of top-ranked Path-VMs in TCGA tasks. Bars show AUROC point estimates for model-task pairs across 19 TCGA tasks. Error bars indicate 95% confidence intervals derived from 1,000 bootstrap replicates, white points indicate AUROC point estimates, and dark red circles denote the top-performing model for each task. Top-, middle-, and lower-ranked panels include n = 95, 95, and 114 model-task AUROC estimates, respectively. e Frequency of top-ranked Path-VMs in non-TCGA tasks. Bars show AUROC point estimates across 22 non-TCGA tasks. Plot elements are defined as in panel d. Top-, middle-, and lower-ranked panels include n = 110, 110, and 132 model-task AUROC estimates, respectively. f Path-VM performance across TCGA tasks. Heatmap shows pairwise AUROC win/loss ratios from Wald tests with bootstrap-estimated variance and FDR correction. The performance score is defined as (wins − losses)/(wins + losses), with darker colors indicating higher net wins; labels denote significant wins/losses at FDR-adjusted p < 0.05. g Path-VM performance across non-TCGA tasks. Heatmap shows pairwise AUROC win/loss ratios across non-TCGA tasks; testing, labels, ranking, and performance score are defined as in (f). AUROC area under the receiver operating characteristic curve, CI confidence interval, FDR false discovery rate, Path-VM pathology-specific vision model, TCGA The Cancer Genome Atlas, CPTAC Clinical Proteomic Tumor Analysis Consortium. Separately, the rankings analysis, which evaluated model performance by averaging median AUROC values across all tasks, identified Virchow2 as the top model (Fig. 1f), followed by Prov-GigaPath in second place, and H-optimus-0 in third. Win/loss tasks, determined through Wald tests assessing statistically significant AUROC differences, were visualized in a heatmap, where wins and losses corroborated the overall rank order (Fig. 1f). Task-level win/loss heatmaps for the top five Path-VMs, derived from Wald tests (FDR-adjusted p < 0.05), provided fine-grained evidence that reinforced the overall ranking patterns (Supplementary Fig. 7a-e, Supplementary Fig. 8a–e) across both TCGA and non-TCGA tasks. Full inferential statistics for the task-level model comparisons underlying the TCGA and non-TCGA heatmaps are provided in Supplementary Data 1 and 2, respectively, including the Wald Z statistic used for testing, the corresponding Wald χ² statistic with 1 degree of freedom, raw p-value, task-wise FDR-adjusted p-value, ΔAUROC, and 95% confidence interval. Performance across all TCGA tasks and non-TCGA tasks for Path-VM models, including mean AUROC values derived from 1000 bootstrap replicates with 95% confidence intervals for each model, is presented in Supplementary Fig. 6 and Supplementary Fig. 9, respectively. Additional performance analyses extended beyond AUROC to include Area Under the Precision-Recall Curve (AUPRC), sensitivity, and specificity, with mean performance revealing Prov-GigaPath leading in AUPRC (Supplementary Fig. 1a), H-optimus-0 in sensitivity (Supplementary Fig. 1c), and UNI2 in specificity (Supplementary Fig. 1e), while on non-TCGA tasks, Virchow2 led in AUPRC, Sensitivity, and Specificity (Supplementary Fig. 1b, 1d, 1f). Task-specific rankings based on these metrics demonstrated Virchow2 leading on 5 out of 19 TCGA-based tasks for AUPRC (Supplementary Fig. 5a), sensitivity (Supplementary Fig. 5c), and specificity each (Supplementary Fig. 5e), UNI in 3 out of 19 tasks for AUPRC, sensitivity and specificity, Prov-GigaPath in 4 out of 19 tasks for AUPRC and specificity, and UNI2 in 3 out of 19 tasks for AUPRC and sensitivity. Across non-TCGA tasks (CPTAC, external benchmarking, and out-of-domain datasets), Virchow2 ranked first in 4 out of 22 tasks for AUPRC (Supplementary Fig. 5b), 7 out of 22 for sensitivity (Supplementary Fig. 5d), and 6 out of 22 for specificity (Supplementary Fig. 5f). Prov-GigaPath achieved top rankings in 2 tasks for AUPRC, 2 for sensitivity, and 1 for specificity, while H-optimus-0 ranked first in 2 tasks for AUPRC, 1 for sensitivity, and 2 for specificity. UNI ranked first in 2 non-TCGA tasks for AUPRC, and UNI2 achieved top rankings in 1 task each for AUPRC, sensitivity, and specificity. Across all TCGA tasks, no statistical differences were observed in AUROC, AUPRC, sensitivity, or specificity among UNI, UNI2, Virchow2, Prov-GigaPath, and H-optimus-0 based on Wald tests, indicating comparable overall performance across metrics (Supplementary Fig. 4a, 4c, 4e, 4g). Across non-TCGA tasks, performance differences were likewise largely non-significant (Supplementary Fig. 4b, 4d, 4h), with the only statistically significant difference observed between Virchow2 and UNI2 for sensitivity (Supplementary Fig. 4f). The corresponding inferential statistics are provided in Supplementary Data 3, including Wald Z, Wald χ²(1), raw p-value, FDR-adjusted p-value, Δmetric, and 95% confidence interval. We further investigated whether specific models excel in tumor-type-specific classification and subtyping tasks, aiming to identify tailored strengths across diverse cancer contexts. Prov-GigaPath demonstrated superior performance on TCGA GBM & LGG, BLCA, and BRCA classification tasks (Supplementary Fig. 2d, Supplementary Fig. 2a–c). In contrast, Virchow2 ranked highest in TCGA PRAD, showcasing its precision in delineating subtle histological subtypes critical for these cancers (Supplementary Fig. 2f). Meanwhile, UNI and UNI2 achieved the highest rankings on lung cancer and pan-cancer assessments respectively. (Supplementary Fig. 2d, e). To statistically evaluate performance differences across tumor types and tasks, we performed pairwise model comparisons using bootstrap-based Wald tests. For each task, AUROC values were estimated from 1000 bootstrap resamples of the evaluation set, and the variance of pairwise AUROC differences was estimated empirically from the bootstrap distribution. Wald statistics were computed using the observed AUROC difference normalized by this variance, and resulting p values were adjusted for multiple comparisons using false discovery rate (FDR) correction. Full inferential statistics for the tumor-type-specific pairwise comparisons are provided in Supplementary Data 4, including Wald Z, corresponding Wald χ²(1), raw p-value, FDR-adjusted p-value, ΔAUROC, and 95% confidence interval. Across tumor types, pairwise Wald tests did not identify statistically significant AUROC differences among UNI, UNI2, Virchow2, Prov-GigaPath, and H-optimus-0 after FDR adjustment (Supplementary Fig. 3a–f; Supplementary Data 5). These results suggest broadly comparable AUROC performance among the top Path-VMs across tumor-type-specific tasks, despite variation in rank order across individual tumor types. Profiling pathology-specific vision model (path-VM) performance across slide and patch-level tasks We evaluated Path-VM across 14 CPTAC and Out of Domain slide-based tasks, including tumor grading, classification, and pathway activity prediction (e.g., p53, MTOR, SWI). Among the 32 models tested, UNI ranked first with a mean AUROC of 0.761 ± 0.083, followed by H-optimus-0 (0.759 ± 0.090). Virchow2 ranked third (0.755 ± 0.089), Prov-GigaPath fourth (0.749 ± 0.091), and H-optimus-mini fifth (0.743 ± 0.092). Across non-TCGA slide-level tasks, we did not observe statistically significant differences in AUROC, AUPRC, sensitivity, or specificity, among the top-performing models, indicating comparable slide-level generalization performance across these models (Fig. 2c, Supplementary Fig. 10a, 10c, 10e). Path-VM performance across non-TCGA slide-level and patch-level tasks. a Path-VM performance in non-TCGA slide tasks. Performance comparison of pathology-specific vision models across non-TCGA slide-level datasets, including CPTAC and out-of-domain tasks. Colored circles represent individual task-level AUROC estimates, with each point corresponding to one model evaluated on one task. Black diamonds indicate the mean AUROC across tasks for each model, and horizontal error bars indicate 95% confidence intervals around the mean, calculated using a t-based standard error across task-level AUROC estimates. For AUROC, n = 14 task-level estimates per model across 16 Path-VMs. b Path-VM performance in patch-based external benchmarking. Performance comparison of pathology-specific vision models across patch-level external benchmarking datasets. Colored circles represent individual task-level AUROC estimates, with each point corresponding to one model evaluated on one task. Black diamonds indicate the mean AUROC across tasks for each model, and horizontal error bars indicate 95% confidence intervals around the mean, calculated using a t-based standard error across task-level AUROC estimates. For AUROC, n = 8 task-level estimates per model across 16 Path-VMs. c Path-VM comparison in non-TCGA slide-level tasks. Box plots summarize task-level AUROC values for Virchow2, UNI, UNI2, Prov-GigaPath, and H-optimus-0 across 14 non-TCGA slide-level tasks. Points represent individual model-task AUROC estimates. Box plots show the median as the center line, interquartile range as box bounds, and whiskers extending to 1.5× the interquartile range; white diamonds indicate the median. Pairwise model comparisons were performed using two-sided bootstrap Wald tests with bootstrap-estimated variance on macro-averaged task performance, with FDR adjustment across model pairs. Exact Wald χ²(1), raw p-values, FDR-adjusted p-values, ΔAUROC, and 95% confidence intervals are reported in Supplementary Data 6. d Path-VM comparison in patch-based external benchmarking tasks. Box plots summarize task-level AUROC values for the same five models across 8 patch-level external benchmarking tasks. Points represent individual model-task AUROC estimates. Box-plot elements and statistical testing are defined as in (c). e Slide-level vs patch-level comparison. Scatter plot comparing mean slide-level AUROC and mean patch-level AUROC for Path-VMs evaluated on non-TCGA datasets. f Heatmap of dataset-specific model performance on non-TCGA datasets. Mean AUROC values for Path-VMs were z-scored across models within each non-TCGA dataset to show relative performance shifts rather than absolute AUROC differences. AUROC area under the receiver operating characteristic curve, CI confidence interval, FDR false discovery rate, Path-VM, athology-specific vision model, CPTAC Clinical Proteomic Tumor Analysis Consortium, OOD out-of-domain. In contrast, on 8 non-TCGA patch-level tasks, differences in mean AUROC were observed, with Virchow2 attaining the highest mean AUROC (0.962 ± 0.033), followed by UNI2 (0.953 ± 0.039), Prov-GigaPath (0.949 ± 0.042), UNI (0.948 ± 0.045), and H-optimus-0 (0.944 ± 0.044)(Fig. 2b); in this setting, some pairwise comparisons reached statistical significance. On these non-TCGA patch-level tasks, pairwise Wald tests identified statistically significant differences in AUROC between Virchow2 and UNI, UNI2, Prov-GigaPath, and H-optimus-0 (all FDR-adjusted p < 0.001), as well as between UNI2 and H-optimus-0 (FDR-adjusted p < 0.01), while the remaining comparisons were not significant (Fig. 2d). Extending this analysis to AUPRC, sensitivity, and specificity for patch-level tasks yielded a similar pattern, with Virchow2 showing statistically significant differences relative to several top models across these metrics, whereas fewer significant differences were observed among the remaining model pairs (Supplementary Fig. 10b, 10d, 10f). Supplementary Data 6 reports the full inferential statistics for these slide- and patch-level comparisons, including Wald Z, corresponding Wald χ²(1), raw p-value, FDR-adjusted p-value, Δmetric, and 95% confidence intervals. Consistent with these results, a direct comparison of slide-level and patch-level performance across non-TCGA tasks showed that models performing well at the slide level generally also achieved strong patch-level performance, although the relative ordering differed modestly between the two settings (Fig. 2e). Finally, a dataset-normalized z-score analysis across CPTAC, external benchmarking, and out-of-domain cohorts highlighted differences in relative model robustness, with top-performing models such as Virchow2, UNI, H-optimus-0, and Prov-GigaPath consistently exhibiting positive standardized performance across multiple non-TCGA settings, while other models showed greater variability across datasets (Fig. 2f). Overall, these findings emphasize that model performance comparisons in computational pathology are highly sensitive to task formulation and evaluation setting, with top-performing models showing comparable slide-level behavior and more variable patch-level performance across non-TCGA benchmarks. Pathology-specific vision models (path-VM) outperform pathology-specific vision–language models (path-VLM) and demonstrate competitive performance relative to general vision models (VM) Next, we systematically evaluated the performance of pathology-specific vision models (Path-VM), pathology-specific vision-language models (Path-VLM; Table 3), and general vision models (VM; Table 4) across classification tasks from TCGA, CPTAC, Out of Domain and eight external benchmarking datasets (Fig. 3a; Supplementary Data 7). Comparative analysis revealed no statistically significant AUROC differences between Path-VM and VM in TCGA, CPTAC, Out of Domain and external benchmarking tasks (Fig. 3a; Supplementary Data 7), with similar non-significant results for AUPRC, sensitivity, and specificity (Supplementary Fig. 11a-c). While Path-VM models frequently occupy top-ranked positions across tasks, we do not observe statistically significant performance differences between Path-VM and VM on TCGA, CPTAC, Out of Domain or external benchmarking datasets; we therefore interpret these results as reflecting greater consistency and robustness rather than uniform statistical superiority. Ranking-based analyses showed that Path-VM models frequently occupied top 10 positions across all datasets, with superior visual feature processing driving this trend (Supplementary Fig. 12a-h, Supplementary Fig. 13, Fig. 3b,c). The pairwise AUROC win/loss heatmaps in Fig. 3b,c were derived from task-level Wald tests with FDR adjustment, with full inferential statistics provided in Supplementary Data 8. In 19 TCGA tasks, Path-VM nearly monopolized the top 10, with rare VM inclusions, such as iBOT achieving third-highest AUROC in bladder and gastrointestinal subtype classification (Supplementary Fig. 14a). For non-TCGA tasks (Fig. 3b, c), Path-VM similarly prevailed, with limited VM success (Supplementary Fig. 15a, b). Notably, select VM instances demonstrated strong task-specific performance, with iBOT and DINO achieving high rankings in lung histology and tumor stage classification on out-of-domain datasets, and DINO model ranking first and fourth on the MGMT methylation and brain necrosis prediction out-of-domain task. In addition, iBOT and DINO-based vision models ranked among the top performers on the MHIST colorectal challenge, further highlighting that general vision models can be competitive on specific non-TCGA tasks (Supplementary Fig. 15b). Model-category performance and pairwise ranking across TCGA and non-TCGA benchmarking tasks. a Path-VM, Path-VLM, and VM performance across TCGA, CPTAC, external benchmarking, and out-of-domain tasks. AUROC distributions are compared between pathology-specific vision models (Path-VM), pathology-specific vision-language models (Path-VLM), and general vision models (VM). Points represent individual model-task AUROC estimates, and white diamonds indicate the mean AUROC for each model category within each dataset. Pairwise comparisons were performed using two-sided Welch two-sample t-tests with FDR adjustment within each dataset and metric. Asterisks indicate FDR-adjusted p-values (* FDR-adjusted p < 0.05, ** FDR-adjusted p < 0.01, *** FDR-adjusted p < 0.001). Exact t statistics with Welch-Satterthwaite degrees of freedom, raw and FDR-adjusted p-values, ΔAUROC, and 95% confidence intervals are reported in Supplementary Data 7. b Path-VM, Path-VLM and VM rankings in non-TCGA Slide-level tasks. Heatmap summarizing pairwise win/loss outcomes among Path-VM, Path-VLM, and VM models based on task-wise comparisons of AUROC. For each task, model pairs were compared using a Wald test with bootstrap-estimated variance across matched replicates, and significance was assessed after FDR correction (FDR < 0.05). The color gradient represents the Performance Score (Wins - Losses)/(Wins + Losses), with darker shades indicating higher net wins. Models are ordered by the number of top-10 task rankings (based on median AUROC per task), with overall rank shown alongside each model. Path-VM’s proportion of 0.47 (106/224) reflects 106 top 10 rankings out of 224 possible slots. c Path-VM, Path-VLM and VM rankings in non-TCGA patch based external benchmarking tasks. Heatmap summarizing pairwise win/loss outcomes among Path-VM, Path-VLM, and VM models based on task-wise comparisons of AUROC. For each task, model pairs were compared using a Wald test with bootstrap-estimated variance across matched replicates, and significance was assessed after FDR correction (FDR < 0.05). The color gradient represents the Performance Score (Wins - Losses)/(Wins + Losses), with darker shades indicating higher net wins. Models are ordered by the number of top-10 task rankings (based on median AUROC per task), with overall rank shown alongside each model. Path-VM’s proportion of 0.52 (66/128) reflects 66 top 10 rankings out of 128 possible slots. AUROC area under the receiver operating characteristic curve, FDR false discovery rate, Path-VM pathology-specific vision model, Path-VLM pathology-specific vision-language model, VM vision model, TCGA The Cancer Genome Atlas, CPTAC Clinical Proteomic Tumor Analysis Consortium. Comparing Path-VM and Path-VLM, Path-VM significantly outperformed Path-VLM across TCGA, CPTAC, and external datasets (Fig. 3a; Supplementary Data 7), with the exception of out-of-domain slide-level tasks. Ranking-based analyses revealed that, despite the overall performance gap, several Path-VLM models achieved competitive, task-specific results. TITAN ranked second on the lung histology OOD2 task and sixth on the CPTAC tumor label task, while both CONCH and TITAN placed within the top five on the UniToPatho challenge. CONCH further demonstrated strong performance on lung-related out-of-domain tasks, ranking first on the lung stage OOD task and within the top three on both lung histology OOD1 and OOD2. In addition, PLIP ranked second on the MGMT methylation out-of-domain task (Supplementary Fig. 17a). In TCGA, a few Path-VLM (BiomedCLIP, QuiltNet-B-16 in bladder subtyping; TITAN in brain histology, PAM50, kidney subtyping) ranked in the top 10 (Supplementary Fig. 16a). Proportionality analysis further showed frequent Path-VM top-10 appearances, with Path-VM occupying 47% of possible positions across datasets (0.55 [168/304] in TCGA, 0.53 [106/224] in CPTAC + OOD, and 0.52 [66/128] in external benchmarking), indicating consistent performance across evaluation settings (Fig. 3b, c, Supplementary Fig. 13; Supplementary Data 8). A broader comparison of vision-only (VM, Path-VM) versus vision-language (VLM, Path-VLM) models showed vision-only models significantly outpacing their language-integrated counterparts across TCGA, CPTAC and external benchmarking tasks (Supplementary Fig. 18; Supplementary Data 9), with Path-VLM’s edge over VLM attributed to pathology-specific text, yet both lagged due to diluted visual learning from multimodal objectives and limited pathology pretraining, emphasizing the critical role of high-resolution visual features in pathology classification. Notably, across out-of-domain slide-level tasks, no statistically significant differences were observed among the four model categories (VM, Path-VM, VLM, Path-VLM) (Supplementary Fig. 18; Supplementary Data 9), indicating that performance advantages observed in TCGA, CPTAC, and external benchmarking datasets do not uniformly translate to fully out-of-domain settings. Collectively, these findings indicate that Path-VM models offer strong and consistent performance across a wide range of pathology classification tasks, characterized by frequent top-ranked placements and a significant advantage over Path-VLM models. At the same time, the absence of statistically significant differences between Path-VM and VM across several evaluation settings suggests that the primary strength of Path-VM lies in robustness and consistency rather than uniform statistical superiority. General vision models demonstrate competitive performance in select tasks, particularly in specific out-of-domain and external challenge-based evaluations, but their more limited presence among top-ranked models highlights the value of pathology-specific training. Path-VLM, trained on histopathology slides and textual data (For example, CONCH36 trained on a multimodal dataset of 1.17 million image-text pairs, and PLIP34 on ~200k image-text pairs), show task specific potential but generally underperformed in our classification benchmarks. Notably, models such as CONCH were first pretrained on large pathology image datasets (e.g., ~16 million image tiles from 21,442 WSIs) before multimodal alignment. Their relative underperformance may therefore reflect limitations in current multimodal architectures or the alignment between image–text supervision and downstream classification tasks, rather than simply differences in dataset scale. These observations suggest that improved multimodal integration may be necessary for Path-VLM models to achieve performance comparable to Path-VM in classification settings. Scaling model size and training data does not uniformly improve pathology model performance We investigated the impact of model and dataset scaling on pathology-specific vision models (Path-VM) across TCGA and non-TCGA datasets (CPTAC, Out of Domain and external benchmarking datasets), evaluating binary and multi-class tasks via AUROC, AUPRC, sensitivity, and specificity. For model scaling, Vision Transformer architectures (ViT-S = 21 M, ViT-B = 85 M, ViT-L > 300 M) were compared across datasets and task-complexity strata using pairwise Welch two-sample t-tests with FDR adjustment, with corresponding inferential statistics reported in Supplementary Data 10. Pairwise effect-size categories were summarized using Cohen’s d (small 0.2, medium 0.5, large 0.8), and Supplementary Fig. 20 visualizes mean performance differences, Cohen’s d effect-size categories, and Spearman correlations between model size, measured in millions of parameters, and performance. Clear scaling benefits were most consistently observed on TCGA tasks, particularly for multi-class settings, where ViT-S to ViT-L comparisons yielded large effect sizes across AUROC, AUPRC, and sensitivity (Supplementary Fig. S20a–d). Spearman correlation analysis further supported these trends, showing moderate and statistically significant associations between model size and performance for TCGA multi-class tasks across multiple metrics (Supplementary Fig. S20e). For non-TCGA datasets, scaling effects were more heterogeneous. CPTAC exhibited modest but statistically significant improvements for binary tasks, particularly for AUROC, AUPRC, and sensitivity when comparing ViT-S to ViT-B and ViT-S to ViT-L(Supplementary Fig. S20a–d; Supplementary Data 10). However, CPTAC multi-class tasks showed no consistent scaling benefit, and correlation analyses did not reveal robust monotonic relationships with model size (Supplementary Fig. S20e). In contrast, external benchmarking and out-of-domain datasets showed weak and non-significant scaling effects across all evaluated metrics, with small effect sizes and negligible correlations between model size and performance (Supplementary Fig. S20a–e; Supplementary Data 10). Across non-TCGA datasets, further scaling from ViT-B to ViT-L generally did not yield statistically significant gains, indicating diminishing returns beyond intermediate model sizes outside TCGA (Supplementary Fig. S20a–d; Supplementary Data 10). Collectively, these findings suggest that model size scaling provides the strongest and most consistent benefits in TCGA, shows limited, task-dependent benefits in CPTAC, and does not reliably improve performance in external or fully out-of-domain settings, underscoring the context-dependent nature of scaling effects in pathology foundation models. We next evaluated the impact of pre-training dataset size by grouping pathology models into small ( < 0.1 M WSIs), medium (0.1–1 M WSIs), and large ( > 1 M WSIs) cohorts. Performance differences were assessed using pairwise effect sizes across AUROC, AUPRC, sensitivity, and specificity (Fig. 4b; Supplementary Fig. S21a–d; Supplementary Data 11), with Spearman correlations used to quantify monotonic associations between dataset size and performance (Supplementary Fig. S21e). TCGA tasks showed the most consistent sensitivity to data scaling. For TCGA multi-class tasks, small-to-medium and small-to-large increases yielded statistically significant improvements in AUROC, AUPRC, and sensitivity (effect sizes ≈ 0.25–0.30, p < 0.05), whereas medium-to-large transitions were uniformly non-significant, indicating diminishing returns beyond moderate dataset sizes (Fig. 4b; Supplementary Fig. S21a–d; Supplementary Data 11). These effects were supported by modest but significant positive correlations between training data size and performance for TCGA multi-class tasks (Supplementary Fig. S21e; Supplementary Data 11). TCGA binary tasks showed weaker and largely non-significant scaling effects across metrics. In contrast, non-TCGA tasks (CPTAC, external benchmarking, and out-of-domain) exhibited limited and inconsistent benefits from increased data scale. CPTAC binary tasks showed modest gains for small-to-medium transitions in AUROC and sensitivity (effect sizes ≈ 0.25–0.40, p < 0.05; Fig. 4b, Supplementary Fig. S21a-d; Supplementary Data 11), but no significant improvements were observed for CPTAC multi-class tasks or for medium-to-large comparisons (Supplementary Fig. S21a–d). External benchmarking and out-of-domain tasks showed uniformly small and non-significant effect sizes across all metrics and data-size comparisons, with weak and non-significant correlations between dataset size and performance (Fig. 4b; Supplementary Fig. S21e; Supplementary Data 11). Specificity remained largely unaffected by dataset scale across both TCGA and non-TCGA settings. Impact of model architecture and pretraining data scale on path-VM performance across benchmarking tasks. a Path-VM comparison of ViT architectures across TCGA, CPTAC, external benchmarking, and out-of-domain tasks. Model performance was evaluated using AUROC across ViT-small, ViT-base, and ViT-large Path-VM architectures. Points represent individual model-task AUROC estimates, with each point corresponding to one model evaluated on one task. Half-eye densities summarize the distribution of task-level AUROC values, and white points indicate the mean with 95% uncertainty intervals. Pairwise architecture comparisons were performed using two-sided Welch two-sample t-tests with FDR adjustment within each dataset and metric. Asterisks indicate FDR-adjusted p-value thresholds (* FDR-adjusted p < 0.05, ** FDR-adjusted p < 0.01, *** FDR-adjusted p < 0.001). Exact t statistics with Welch-Satterthwaite degrees of freedom, raw p-values, FDR-adjusted p-values, ΔAUROC, and 95% confidence intervals are reported in Supplementary Data 10. b Path-VM performance by pretraining data scale across TCGA, CPTAC, external benchmarking, and out-of-domain tasks. Model performance was evaluated using AUROC as a function of the number of whole-slide images used during pretraining. Pretraining dataset sizes were grouped as small ( < 0.1 million slides), medium (0.1–1 million slides), and large ( ≥ 1 million slides). Points represent individual model-task AUROC estimates, with each point corresponding to one model evaluated on one task. Half-eye densities summarize the distribution of task-level AUROC values, and white points indicate the mean with 95% uncertainty intervals. Pairwise comparisons across pretraining data-scale groups were performed using two-sided Welch two-sample t-tests with FDR adjustment within each dataset and metric. Asterisks indicate FDR-adjusted p-value thresholds (*FDR-adjusted p < 0.05, **FDR-adjusted p < 0.01, ***FDR-adjusted p < 0.001). Exact t statistics with Welch-Satterthwaite degrees of freedom, raw p-values, FDR-adjusted p-values, ΔAUROC, and 95% confidence intervals are reported in the Supplementary Data 11. AUROC area under the receiver operating characteristic curve, CI confidence interval, FDR false discovery rate, Path-VM pathology-specific vision model, ViT vision transformer, TCGA The Cancer Genome Atlas, CPTAC Clinical Proteomic Tumor Analysis Consortium. Collectively, these results indicate that increasing pre-training dataset size yields measurable benefits primarily for TCGA tasks, with more limited and task-dependent effects in non-TCGA settings (Fig. 4b). The diminishing returns observed for external and out-of-domain evaluations suggest that scaling alone may be insufficient to support robust generalization, motivating further investigation into how dataset characteristics influence transfer performance. Overall, these findings indicate that neither increasing model size nor pre-training dataset size uniformly guarantees improved performance in pathology foundation models. Scaling benefits are most consistently observed for TCGA tasks, with additional gains in select CPTAC settings, whereas performance improvements are more limited and task-dependent in external and out-of-domain evaluations. In particular, large architectural jumps (ViT-S to ViT-L) and moderate increases in training data yield measurable benefits, while intermediate scaling often results in diminishing returns. We note that these scaling analyses do not explicitly account for differences in pre-training data composition or domain similarity across models, which may contribute to the observed variability in task-specific generalization performance. Model ensembling improves cancer task performance with variable gains across tissues and domains To explore whether combining models could improve predictive performance, we implemented a model ensembling approach using a late decision-level fusion strategy (referred to hereafter as the fusion model) integrating predictions from the top five models identified from non-TCGA rankings (Virchow2, H-optimus-0, Prov-GigaPath, UNI, and UNI2). In this fusion model, predictions from these top-performing foundation models were aggregated using a non-trainable majority-vote scheme. Models were ranked based on cumulative balanced accuracy, separately for TCGA tasks and for the combined set of CPTAC, out-of-domain, and external benchmarking tasks (non-TCGA). Across TCGA tasks, the fusion model achieved the highest mean cumulative balanced accuracy when aggregated over 19 tasks (0.702 ± 0.111), exceeding the best individual model, Virchow2 (0.696 ± 0.105), as well as other top-performing models (Fig. 5a). However, pairwise Wald tests at the individual task level did not identify statistically significant differences between the fusion model and any single top-performing model (Virchow2, Prov-GigaPath, H-optimus-0, UNI, or UNI2), indicating that the fusion model’s advantage reflects aggregate performance rather than consistent per-task superiority (Fig. 5c; Supplementary Data 12). Across non-TCGA tasks (CPTAC, external benchmarks, and out-of-domain), Virchow2 achieved the highest mean balanced accuracy (0.712 ± 0.200), followed by the fusion model (0.699 ± 0.203) and UNI (0.698 ± 0.179) (Fig. 5b). Trends in cumulative sensitivity mirrored those observed for balanced accuracy (Supplementary Fig. 22a–b). At the task level, pairwise win–loss comparisons favored the fusion model over UNI (2–0), H-optimus-0 (3–0), Prov-GigaPath (2–0), and UNI2 (1–0)(Fig. 5d); however, these trends did not translate into consistent statistical significance across tasks (Supplementary Fig. 23a-b; Supplementary Data 13). Model ranking, fusion performance, and tissue-level specialization across benchmarking tasks. a Model rankings in TCGA tasks. Stacked bars show cumulative balanced accuracy across 19 TCGA tasks, with each colored segment representing one task-level balanced accuracy estimate. Error bars indicate propagated 95% confidence intervals for cumulative balanced accuracy, calculated by combining task-level 95% confidence interval half-widths. Models are ordered by cumulative balanced accuracy; the fusion model ranks first. b Model rankings in non-TCGA tasks. Stacked bars show cumulative balanced accuracy across 22 CPTAC, external benchmarking, and out-of-domain tasks. Bar segments and error bars are defined as in panel a. Models are ordered by cumulative balanced accuracy; the fusion model ranks second. c TCGA pairwise model-ranking heatmap. Heatmap summarizes task-wise pairwise comparisons of balanced accuracy across TCGA tasks. For each task, model pairs were compared using Wald tests with bootstrap-estimated variance across matched replicates, followed by FDR correction. Cells report significant wins/losses for the row model versus the column model. The performance score is defined as (wins − losses)/(wins + losses), with darker colors indicating higher net wins. d Non-TCGA pairwise model-ranking heatmap. Heatmap summarizes task-wise pairwise comparisons of balanced accuracy across CPTAC, external benchmarking, and out-of-domain tasks. Statistical testing, FDR correction, cell labels, and performance score are defined as in (c). e Model performance across all tasks. Heatmap shows AUROC across classification tasks. For each task, the top 10 models were selected by mean AUROC across test replicates, and performance was z-scored within the task; blank cells indicate models outside the top 10. Side annotations indicate model category. f Model performance across tissues. Heatmap shows tissue-specific AUROC. For each tissue, the top 10 models were selected by mean AUROC across test replicates, and performance was z-scored within the tissue. Numerical annotations indicate mean AUROC, and side annotations indicate model category. AUROC area under the receiver operating characteristic curve, CI confidence interval, FDR false discovery rate, Path-VM pathology-specific vision model, Path-VLM pathology-specific vision-language model, VM vision model, TCGA The Cancer Genome Atlas, CPTAC Clinical Proteomic Tumor Analysis Consortium. Individual models also exhibited shifts in relative ranking between TCGA and non-TCGA settings (Fig. 5a-b), reflecting differences in cross-domain generalization. UNI improved from 5th on TCGA tasks (0.693 ± 0.111) to 3rd on non-TCGA tasks (0.698 ± 0.179), whereas UNI2 declined from 6th (0.684 ± 0.117) to 9th (0.682 ± 0.195). Virchow similarly shifted from 7th on TCGA (0.677 ± 0.117) to 10th on non-TCGA tasks (0.678 ± 0.197). These rank changes highlight dataset-dependent robustness rather than a fixed performance hierarchy. Consistent with these observations, pairwise Wald tests did not identify statistically significant differences between the fusion model and individual top-performing models across either TCGA or non-TCGA tasks (Supplementary Fig. 23a–b; Supplementary Data 12 and 13). To examine tissue-specific specialization, we computed z-scored AUROC values for each model and grouped the 41 tasks by tissue type. Rather than a single model dominating all tissues, we observed complementary performance profiles across models (Fig. 5e-f). CONCH achieved the strongest performance in lung-related tasks (Fig. 5f), Virchow2 led in colon and prostate while remaining competitive in lung, and H-optimus-0 performed best in pan-cancer tasks (Fig. 5f). UNI showed strong performance in brain and bladder, UNI2 was competitive in breast and pan-cancer, and Prov-GigaPath displayed robust performance across breast, bladder, and lung (Fig. 5f). EXAONEPath also showed strong performance in brain tasks, ranking among the top models in this tissue group (Fig. 5f). Pairwise Wald comparisons based on AUROC indicated that Virchow2 was significantly better than the remaining top models in colon tissue, while UNI2 outperformed H-optimus-0 and UNI in colon, with no significant differences observed in other tissues (Supplementary Fig. 24a–f; Supplementary Data 14). Extending this analysis to balanced accuracy, the fusion model achieved the highest performance in breast and pan-cancer tasks (Supplementary Fig. S22c–d) and remained consistently competitive in bladder, brain, and prostate, frequently ranking among the top models. Statistical comparisons showed that the fusion model was significantly better than UNI, UNI2, and H-optimus-0 in colon tissue, while remaining statistically indistinguishable from Virchow2 and Prov-GigaPath and showing no significant differences in other tissues (Supplementary Fig. 23c–i; Supplementary Data 15). Finally, when grouping tasks into biomarker status prediction, diagnostic classification, gene expression subtype prediction, and pathway activity prediction, no statistically significant differences were observed among models for balanced accuracy (Supplementary Fig. 25a–h; Supplementary Data 15). Together, these findings demonstrate that model fusion achieves the strongest cumulative performance on TCGA tasks while maintaining competitive generalizability across non-TCGA settings. Although Virchow2 remained the strongest individual performer on out-of-domain tasks, the fusion model generally ranked ahead of the remaining top TCGA models in non-TCGA evaluations, highlighting its robustness beyond the training domain. Rather than uniformly dominating across all settings, the fusion model reflects complementary strengths across tissues and task categories, resulting in stable and reliable performance across cancer types. Discussion This study presents a large-scale benchmarking analysis of 32 foundation models spanning general vision (VM), general vision-language (VLM), pathology-specific vision (Path-VM), and pathology-specific vision-language (Path-VLM) architectures across 41 pathology tasks derived from TCGA, CPTAC, external benchmarking datasets, and institution-specific cohorts. All models were evaluated using a standardized feature-extraction and linear-probing pipeline to ensure fair comparison across architectures and training paradigms. The goal of this study was to investigate the task-specific strengths of various models and assess their overall performance on common cancer-specific tasks. Our findings reveal notable variations in model performance, reflecting the inherent complexity of pathology tasks and emphasizing the necessity for tailored approaches depending on the type of cancer and the specific challenge. In our study, it is crucial to also acknowledge that TCGA (The Cancer Genome Atlas)30 likely serves as a major source of training data for many of the models evaluated. Although self-supervised learning reduces reliance on labeled data, exposure to TCGA-derived distributions during pretraining cannot be ruled out for all models. A central aim of this work was to characterize task- and domain-specific performance patterns while explicitly distinguishing in-domain (TCGA) evaluation from non-TCGA generalization, which includes CPTAC, external patch-level benchmarks, and out-of-domain slide-level cohorts. This separation is critical given that TCGA likely contributes, directly or indirectly, to the pretraining data of several evaluated models25,33,54, and therefore primarily reflects in-domain performance rather than strict generalization. Across TCGA tasks, multiple pathology-specific vision models demonstrated strong and statistically comparable performance, with Virchow2, Prov-GigaPath, H-optimus-0, UNI, and UNI2 consistently occupying the top ranks. Within this group, models exhibited distinct tissue-specific strengths25,28,55: Virchow2 performed particularly well in colon,lung and prostate tasks, Prov-GigaPath showed strong performance in breast and bladder, H-optimus-0 excelled in pan-cancer tasks, and UNI demonstrated competitive performance in brain, bladder, and breast-related tasks. Extending evaluation to non-TCGA datasets revealed more nuanced generalization behavior. While several models maintained competitive performance across CPTAC and external benchmarks, relative rankings showed modest but consistent shifts, underscoring that strong TCGA performance does not uniformly translate to out-of-domain robustness. In particular, Virchow2 remained the strongest individual performer across non-TCGA evaluations, whereas other top TCGA models exhibited dataset-dependent rank changes rather than a fixed performance hierarchy. These findings emphasize that cross-domain generalization in computational pathology is heterogeneous, tissue dependent, and highly task specific. A key contribution of this work is the explicit comparison between slide-level and patch-level evaluation settings across non-TCGA datasets. While slide-level tasks (TCGA, CPTAC, and out-of-domain cohorts) showed largely comparable performance among top-performing models, patch-level external benchmarks revealed clearer performance separation, with statistically significant differences emerging for several model pairs. This divergence suggests that patch-level tasks, which lack multi-instance aggregation, may more directly probe fine-grained visual representations, whereas slide-level performance reflects both feature quality and aggregation effects56. Importantly, models that performed strongly at the slide level generally remained competitive at the patch level, although relative ordering varied, highlighting that evaluation granularity substantially influences apparent model performance. When comparing model categories, Path-VM models consistently achieved top-ranked placements across both TCGA and non-TCGA tasks, reflecting robust representation learning tailored to histopathology. However, we did not observe statistically significant differences between Path-VM and general VM models across most evaluation settings, including TCGA, CPTAC, and external benchmarks. As such, we interpret the advantage of Path-VM primarily in terms of consistency and robustness, rather than uniform statistical superiority. In contrast, Path-VLM models underperformed relative to Path-VM across most tasks, likely due to limited availability of large-scale, well-curated pathology image–text datasets and challenges in effectively integrating high-resolution visual information with textual supervision. While select Path-VLMs demonstrated strong task-specific performance, particularly in lung-related and challenge-based evaluations, further advances in multimodal training strategies are required to close the performance gap. Our scaling analyses further demonstrate that neither increasing model size nor expanding pretraining dataset size guarantees improved performance. Scaling benefits were most consistently observed for TCGA tasks, particularly for multi-class settings, whereas CPTAC exhibited modest, task-dependent gains and external or fully out-of-domain datasets showed little sensitivity to scaling. Both architectural scaling and dataset expansion exhibited diminishing returns beyond moderate sizes. Importantly, these analyses are correlational and do not account for differences in pretraining data composition, tumor-type prevalence, or sampling strategies, which likely contribute to observed task-specific performance differences. Where publicly available, differences in tissue-specific performance appear broadly consistent with heterogeneity in pretraining data composition across top performing models (Supplementary Tables 3-8). For example, Virchow2 was trained on large and diverse pathology cohorts with substantial representation from breast, colon, and prostate tissues, and correspondingly demonstrated strong performance across these tissue groups in downstream evaluations. Prov-GigaPath, whose pretraining data are heavily enriched for lung and gastrointestinal slides, showed robust performance in lung tasks and competitive results in breast and bladder, suggesting that its representations transfer effectively beyond its most prevalent training tissues. UNI was trained on a wide range of tissue types, including brain, bladder, and breast, and exhibited strong or competitive performance across these tissue groups in both TCGA and non-TCGA settings. Similarly, CONCH demonstrated its strongest relative performance in lung-related tasks despite having a smaller absolute number of lung images compared with some vision-only models, but a comparatively high proportion of lung tissue within its multimodal pretraining corpus (SuppTable 8), suggesting that relative tissue composition rather than dataset scale alone may influence downstream generalization. For other top-performing models, including UNI2 and H-optimus-0, detailed tumor-type composition of pretraining datasets is not publicly disclosed, limiting our ability to assess whether observed tissue-specific strengths reflect training data composition, architectural choices, or other factors. Importantly, these observations are descriptive and do not imply a causal relationship between tissue prevalence during pretraining and downstream task performance, as multiple interacting factors are likely to influence model generalization. Model ensembling provided an additional perspective on robustness. A late fusion model constructed from the top-performing non-TCGA models achieved the highest cumulative performance on TCGA tasks and maintained competitive rankings across non-TCGA evaluations. However, these gains reflected aggregate trends rather than consistent per-task superiority, with pairwise statistical tests showing few significant differences relative to individual models. Tissue-level analyses further indicated that fusion benefits were concentrated in select tissues (notably breast and pan-cancer), while performance remained statistically comparable to top individual models in most other settings. These results suggest that fusion models effectively aggregate complementary strengths without universally outperforming strong individual encoders. Given that ensemble methods are often more resilient to unseen data57–59, fusion models offer a compelling strategy for improving diagnostic AI in pathology. Several limitations should be noted. First, our evaluation relies on frozen feature extraction with linear probing, which benchmarks representation quality but does not reflect the full performance potential under task-specific fine-tuning. Second, potential overlap between public datasets (e.g., TCGA, CPTAC) and model pretraining data cannot be fully excluded and may inflate in-domain performance estimates. Third, our scaling analyses use coarse proxies such as parameter count and number of slides, which do not capture critical aspects of data composition or quality. Finally, slide-level aggregation relies on simple mean pooling, which may attenuate localized discriminative signals. Our findings are broadly consistent with recent pathology foundation model benchmarking studies50,51,60. These studies reported strong performance from newer pathology-specific foundation models, including Virchow2, Prov-GigaPath, H-optimus-0, UNI, while older pathology encoders such as CTransPath generally underperformed. A recent benchmark60 across lung, breast, gastric, colorectal, and brain cancers similarly found that vision foundation models such as Virchow2 and H-optimus models were more effective than vision-language models for clinical-level tasks, with Virchow2 ranking best in breast, brain, and gastric cancer tasks. Our results support these same trends across TCGA, CPTAC, and external datasets. Across studies, model size and pretraining dataset size alone were not reliable predictors of downstream performance; instead, model performance appears tissue- and task-dependent, with specific models excelling in particular tumor types or clinical endpoints. These observations highlight the importance of tissue diversity, pretraining composition, and training strategy, consistent with prior reports that tissue-specific pretraining effects, particularly for lung-focused biomarker tasks, may contribute to downstream strengths. In addition, our ensemble-based analyses suggest that combining complementary pathology foundation models can improve robustness and generalization across datasets and tissue types. In contrast to prior benchmarking efforts50,51 that evaluated a limited number of tissues or task types, our study evaluates foundation models across a broad set of histopathologic tasks spanning multiple tissue types, datasets, and task categories, enabling a more comprehensive assessment of model generalization. In addition, our benchmark systematically compares general vision, vision-language, pathology-specific vision, and pathology-specific vision-language models under a unified evaluation framework. We also evaluate recently released models, including UNI2, H-optimus-mini, and EXAONEPath, allowing the benchmark to reflect the rapidly evolving landscape of pathology foundation models. Furthermore, our task set includes pan-cancer evaluations, which assess whether model representations generalize across multiple tumor types rather than only within tissue-specific settings. Overall, this benchmark suggests that model size and pretraining dataset scale alone do not consistently explain performance differences in computational pathology tasks. Although we analyzed these factors explicitly, additional variables such as tumor type prevalence, dataset composition, data diversity, tile sampling strategies, and acquisition or preprocessing characteristics are also likely to influence downstream performance. By explicitly separating TCGA and non-TCGA evaluations, disentangling slide- and patch-level behavior, and contextualizing model ensembling and scaling effects, our study provides a more nuanced view of model generalization and practical deployment considerations in pathology. The results highlight that, while generalization is a key challenge, task-specific model optimization and ensemble-based approaches can meaningfully improve model performance. Future work may further explore the potential of multimodal models that integrate both image and textual data, as well as strategies addressing data diversity16,61 and optimization techniques for pathology-specific vision models (Path-VM), to enhance model transferability across diverse datasets. Methods Ethics statement This study complied with all relevant ethical regulations. Publicly available datasets were used in accordance with the terms and conditions specified by the respective data providers. The Stanford non-small cell lung cancer dataset used for out-of-domain evaluation was derived from a cohort approved by the Stanford Institutional Review Board (IRB #48496), and Stanford GBM dataset was approved by Stanford Institutional Review Board (IRB #50031). The requirement for informed consent was waived by the IRB because of the deidentified retrospective nature of the study. The Stanford dataset was accessed only by authorized Stanford research personnel under applicable institutional approvals and data-governance requirements. Sex- and gender-stratified analyses were not performed because sex and gender metadata were not consistently available across the benchmarking datasets and were not used as model inputs. The study focused on model performance across histopathology tasks, datasets, model categories, and tissue/task types rather than participant-level demographic subgroup analysis. Whole slide image preprocessing Whole-slide images H&E (WSIs), initially in the SVS format, needed to be preprocessed to a machine learning-ready format. Due to the massive size of WSIs, which can often exceed dimensions of 10,000 ×10,000 pixels, it was necessary to divide them into smaller, non-overlapping tiles of 256 ×256 pixels each, at an objective magnification of 20x (0.5 microns per pixel). To accomplish this, a tissue mask was first obtained using the Otsu thresholding method62, which effectively separates the foreground (tissue) from the white background. A multiple instance learning (MIL) strategy was employed, where each WSI was represented by a selection of up to 4,000 tiles. This approach was adopted to address the challenge of variable tissue density and distribution within WSIs. Tiles with excessive background or low contrast were excluded to ensure the selection of informative and representative tiles for each WSI. The selected tiles were then saved in an HDF5 database, one database per slide, to facilitate fast input/output (I/O) operations during the training and inference stages. This database organization allows for efficient access to the tiles, reducing the computational overhead associated with loading and processing large WSIs. During the training phase, tiles were grouped into bags of G size. Then, in each bag, a feature vector was extracted for each tile using a given foundation model. A maximum of G tiles was considered in each bag, and the corresponding feature vectors were averaged. This averaged feature vector was then forwarded through a linear layer to obtain the final classification result for each WSI. This approach effectively leverages the MIL framework to model the variability within WSIs and enables the classification of slides based on the collective information extracted from their constituent tiles. The use of the HDF5 database and the efficient tile-based representation contribute to the scalability and computational efficiency of the overall pipeline. Whole slide image preprocessing was conducted for the TCGA, CPTAC, and Out-of-Domain datasets, while no additional preprocessing was required for External benchmarking tasks, as the images were conveniently provided in a ready-to-use.png format. Vision transformers Vision Transformers (ViTs) have significantly advanced SSL in the realm of computer vision. Originally designed for natural language processing, ViTs adapt the transformer architecture to visual tasks by dividing images into smaller patches treated as tokens. The self-attention mechanisms in ViTs enable the model to capture intricate details and long-range dependencies within images, providing a superior contextual understanding compared to traditional convolutional neural network-based methods, especially in complex domains like pathology63. Foundation models In the past couple of years, several ViT based pathology models like Lunit-Dino24, Phikon(Owkin)33, UNI25, Virchow26 have been trained on huge datasets on multiple tissue types and evaluated on a variety of downstream tasks. We evaluated in-domain foundation models (i.e., models trained on digital pathology data such as Lunit-Dino), as well as out-of-domain vision models, trained on natural images (ImageNet data) such as DINO from Meta AI. We further categorized these models as general vision only (natural image based), general vision-language (natural images with text), pathology vision models (digital pathology image-based) and pathology vision-language models (pathology images and medical text). We conducted a comparative analysis across the different model categories to assess their relative performance. The pathology domain-specific models have been primarily trained on publicly available datasets like TCGA, and/or on proprietary datasets (for example, in the case of UNI). These foundation models are built upon ViT-small, ViT-base, and ViT-large/ViT-huge architectures as their core backbones. In our work, we benchmark these models using identical tasks, datasets, and hyperparameters. Computational resources For our evaluation experiments, we used the Polaris supercomputer at the Argonne National Laboratory and the Stanford University Sherlock cluster (Stanford Research Computing Center) as well as Stanford University Marlowe cluster64. Model evaluation We assessed the performance of 32 foundation models using a standardized linear-probing protocol applied to frozen feature representations. For all downstream classification tasks, whole-slide images were first sampled at 20x magnification, and image patches of size 256 × 256 pixels were extracted to define a consistent physical tissue context across datasets. Prior to feature extraction, these patches were resized to the input resolution required by each model, which was 224 × 224 pixels for the majority of evaluated encoders, following their native preprocessing pipelines. All pretrained models were kept frozen during evaluation. Feature embeddings were extracted once per task and used as input to a linear classifier trained on top of the frozen representations. Linear probing was performed using a learning rate of 0.001, a batch size of 16 per GPU, and training for 30 epochs, without data augmentation. This protocol was applied uniformly across all models to enable a controlled comparison of representation quality under consistent evaluation settings. Experiments were conducted on a single node equipped with four NVIDIA A100 GPUs (40 GB each). The foundation models differed in architecture (e.g., ViT-S, ViT-B, ViT-L) and patch size. For downstream tasks using TCGA, CPTAC, and out-of-domain datasets, we applied a 70/30 split (patient-level split) into training, and hold-out test sets. To enhance robustness, we also performed bootstrapping with 1000 replicates to calculate 95% confidence intervals. We report the performance on these independent test sets throughout our study. For external public benchmarking datasets, where test or validation sets are predefined, we report results on those respective splits. We evaluated models on both slide-level tasks from TCGA, CPTAC, and out of domain datasets, as well as patch-level classification tasks from independent external benchmarks. TCGA, CPTAC and Out of Domain evaluations are performed as slide-level classification tasks, in which predictions are generated by aggregating features across multiple tiles extracted from each whole-slide image. In contrast, the external benchmarking datasets consist of tile-level (single-image) classification tasks, where each image is evaluated independently without tile aggregation. All performance metrics were computed on distinct, non-overlapping samples. That is, measurements were taken from distinct samples only, with no reuse across training or test sets. This ensured a clean separation between evaluation conditions and avoided data leakage. Model effectiveness was quantified using AUROC, AUPRC, Sensitivity and Specificity. Balanced accuracy was used to evaluate the fusion model with other models, which ensembled top-performing models through a majority voting strategy. Statistical analysis of model comparisons To compare the performance of foundation models across cancer types, task categories and tissues, we evaluated area under the receiver operating characteristic curve (AUROC) distributions. Model performance was benchmarked across TCGA, CPTAC, and External benchmarking cohort datasets, aggregating replicate-level results from multiple models. Model performance estimates were obtained using 1000 bootstrap replicates of the evaluation data, providing stable estimates of AUROC distributions and their variance. We have implemented the bootstrap method to make inferences about the model performance in the test set. Specifically, the test set was bootstrapped 1000 times and AUROCs from all prediction models were calculated for each bootstrapped test set. The 95% confidence interval of the AUROC for a given model was defined by the 2.5th and 97.5 percentiles of 1000 bootstrapped AUROC estimates. To compare two models, we also calculated the difference in AUROC between these two models in each bootstrapped test set. The variance of the estimated difference in AUROC was then estimated by the empirical variance of the 1000 bootstrapped AUROC differences. A Wald test was then performed using the ratio of the observed mean AUROC difference to the estimated standard deviation, yielding a two-sided p-value under a normal approximation. p-values were adjusted for multiple comparisons using False Discovery Rate (FDR) method, establishing significance at an FDR-adjusted p-value threshold of <0.05 to control for false positives. Task-level performance was summarized through aggregated win-loss matrices, where a “win” was recorded if a model demonstrated significantly higher AUROC across tasks, and a “loss” if significantly lower, based on the FDR criterion. Net performance scores were calculated as (wins - losses)/(wins + losses + 1), providing a normalized measure of relative dominance ranging from -1 to 1. Model rankings were derived by computing the median AUROC per task, averaging these values across all tasks, and ordering models by their mean median AUROC, with normalized rank scores scaled from 0 to 1 based on rank position to facilitate cross-model comparisons. For comparisons involving the fusion model, balanced accuracy was used as the performance metric to account for the majority-voting scheme employed to aggregate predictions from the top-performing models. Proportionality analysis assessed the frequency of top 10 rankings per model category (Path-VM, Path-VLM, VM). For each task and metric, models were ranked by median performance, and those in the top 10 were counted. The proportion was calculated as the number of top 10 appearances (n_top10) divided by the total possible top 10 slots (number of models per category × number of tasks). For example, In CPTAC and out of domain tasks, Path-VM’s proportion of 0.47(106/224) reflects 106 top 10 rankings out of 224 possible slots. Performance was assessed using two approaches: heatmap analysis summarized pairwise net performance scores and win-loss ratios across tasks, with diagonal elements set to zero; boxplot analysis visualized AUROC distributions per task, highlighting medians and annotating significant and meaningful pairwise comparisons (p < 0.01, p < 0.001), with no annotations for non-significant comparisons. Both analyses ordered models by average median AUROC to reflect relative performance. Statistical analysis of model and data size Model performance was assessed across TCGA, CPTAC, and External benchmarking cohort datasets, focusing on pathology-specific vision models (Path-VM) with varying architectural scales (small, base, large). Tasks were categorized into binary and multi-class complexity based on their classification structure to evaluate performance under different levels of difficulty. Performance metrics, including Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), sensitivity, specificity, weighted F1 score, and balanced accuracy, were computed as averages across 1000 bootstrap replicates to account for variability and provide stable estimates with 95% confidence intervals. For each metric, data were stratified by dataset and task complexity, and comparative analyses were conducted across model architectures. Statistical significance of performance differences between model sizes (small vs. base, base vs. large, small vs. large) was evaluated using two-sided t-tests, with p-values adjusted for multiple comparisons using the False Discovery Rate (FDR) method, setting a significance threshold of p < 0.01. These comparisons were visualized to illustrate distribution patterns and highlight significant variations, facilitating the assessment of scaling effects on model efficacy across diverse task complexities and datasets. To further probe scaling impacts, a contrast-driven analysis quantified mean performance differences between architectural scales (small vs. base, base vs. large, small vs. large) for each metric, normalized to a [-1, 1] scale to standardize variation. Effect sizes, expressed as Cohen’s d, were calculated by dividing the t-test statistic by the square root of the sample size, offering a standardized measure of the magnitude of performance differences. Cohen’s d was interpreted with established thresholds—small (0.2), medium (0.5), and large (0.8)—to assess practical significance. Results were visualized in heatmaps, integrating normalized differences and Cohen’s d effect sizes, annotated to reveal the differential scaling benefits across datasets and task complexities. Spearman correlation analyses were visualized in scatter plots, annotated with rho and p-values, to evaluate the relationship between model size and performance. Software Model evaluation and feature extraction were performed using Python 3.8.20 in a conda environment provided in the accompanying GitHub repository. Key packages included PyTorch 2.3.1, torchvision 0.18.1, timm 1.0.11, transformers 4.39.2, scikit-learn 1.3.2, scipy 1.10.1, pandas 2.0.3, NumPy 1.24.4, h5py 3.11.0, OpenCV 4.10.0.84, scikit-image 0.21.0, Pillow 10.4.0, and Hugging Face Hub 0.26.1. Statistical analysis and figure generation were performed in R, with plotting and statistical packages listed in the accompanying code repository. The full conda environment file and analysis scripts are available at https://github.com/gevaertlab/benchmarking-path-models Datasets for downstream tasks TCGA dataset The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/) project aimed at cataloging and generating comprehensive molecular landscapes responsible for cancer. The Cancer Genome Atlas project30 generated multiple omics modalities, including exome-seq, RNA-seq, and methylation data, along with digital pathology slides, covering over 32 tumor types and 10,000 patients. With respect to digital pathology, TCGA provided a vast collection of high-resolution whole-slide images of tumor sections stained with Hematoxylin and Eosin (H&E). This has facilitated the development of machine learning and deep learning algorithms for tumor subtype classification, prognostication, and treatment response prediction using pathology imaging data. We used this rich resource of data for evaluation of foundation models. We evaluated the models based on tumor-specific molecular subtyping, such as distinguishing between IDH mutant and wild-type brain tumors, as well as pan-cancer molecular subtyping, like differentiating between TP53 mutant and wild-type tumors. This approach provided an estimate of the models’ performance in specific tumors and across a pan-cancer setting. Additionally, we evaluated these models on a diverse range of tasks, including pathway activity, immune group subtyping, mRNA cluster classification, and histological subtyping, across various cancer types. Overall, we curated around 19 downstream tasks from TCGA data, encompassing bladder, brain, prostate tumors, and more. CPTAC The Clinical Proteomic Tumor Analysis Consortium65 (CPTAC) has provided pan-cancer datasets that include omics and digital pathology data, similar to the TCGA project. We also used this resource (https://proteomics.cancer.gov/data-portal) to evaluate our models. Since most pathology models have been trained on the TCGA dataset, testing their performance on CPTAC pathology data will offer a true measure of their efficacy on an independent test dataset. From CPTAC, we curated 7 pan-cancer tasks66: grade, SWI pathway alteration, p53, chromatin modification pathway alteration, mTOR pathway alteration, MYC pathway alteration, tumor label. Out of domain datasets We used proprietary pathology datasets from Stanford, including lung and brain tumor samples. Given that some foundation models may have used CPTAC data during their pre-training, it was crucial to evaluate these models on our in-house datasets to test their generalizability and estimate their performance on real-world data. With access to detailed clinical information and molecular subtypes, we assessed these models on specific subtyping tasks, such as distinguishing between lung squamous cell carcinoma and adenocarcinoma, and differentiating between MGMT methylated and wild-type brain tumors. We also added NLST pathology67 and DHMC lung cancer datasets68 to this category as a part of our out-of-domain evaluation. This approach helped us determine the models’ true effectiveness on practical clinical tasks. In this study, the term out-of-domain (OOD) refers to datasets that differ from TCGA and CPTAC in acquisition site, preprocessing pipelines, and task definitions, rather than implying guaranteed exclusion from model pretraining. While NLST and DHMC are not reported as part of the training datasets for the evaluated models, this assumption cannot be independently verified and therefore represents a limitation. External benchmarking datasets We curated 8 publicly available external benchmarking datasets, including BRACS69, BACH70, UnitoPatho71, SICAPv272, BreakHis73, LC250074, MHIST75, and NCT-CRC-HE76. These datasets contain representative images from diverse tumor types and are completely independent of the training data, ensuring a robust and unbiased evaluation of the foundation models. For the BreakHis challenge dataset, we used a single fold for both training and evaluation. All external benchmarking datasets involved patch-level classification tasks. Additional details are provided in Supplementary Table 1. In this table, the reported numbers for external benchmarking tasks correspond to the number of image patches, whereas for TCGA, CPTAC, and out-of-domain datasets, N denotes the number of unique patient slides. Slide-level and patch-level evaluation The downstream tasks included both slide-level and patch-level classification settings. For TCGA, CPTAC, and out-of-domain datasets, predictions were performed at the slide level. Whole-slide images were first divided into non-overlapping image patches, which were processed by the foundation model to extract patch-level feature embeddings. These patch-level representations were then aggregated into slide-level features, enabling classification at the whole-slide level. In contrast, the external benchmarking datasets consisted of curated image patches and were therefore evaluated directly using patch-level classification. In this setting, each patch was treated as an independent sample and predictions were computed directly from patch-level representations without slide-level aggregation. Separating slide-level and patch-level evaluation allows assessment of model behavior across different levels of histopathologic context and reflects the varying data structures commonly encountered in computational pathology benchmarks. Metadata collection for downstream tasks For the TCGA tasks, we collected molecular subtype, gene mutation status, histologic subtype, and transcriptomic information from TCGA publications specific to the following cancer types: brain77, prostate78, lung79,80, breast81, and bladder82. For the pan-cancer tasks, we collected pathway information on the samples from the TCGA pan-cancer pathway study83 and immune groups from the TCGA immune landscape study84. For the CPTAC tasks, we gathered molecular subtypes and activated pathways on cases from the CPTAC pan-cancer study66. Downstream tasks The downstream tasks included H&E stained whole slide-level using TCGA, CPTAC, and out-of-domain datasets, with a 70/30 split into training, and evaluation on independent held-out test sets. For external benchmarking datasets (patch-level classification) with predefined splits, performance was reported on those sets, using AUROC, AUPRC, Sensitivity and Specificity on non-overlapping samples. Model evaluation followed, applying a linear classifier to frozen features extracted from H&E 224 × 224 pixel images at ×20 magnification, with probing conducted over 30 epochs using a 0.001 learning rate and 16 batch size per GPU on four NVIDIA A100 GPUs. Supplementary Table 1 provides the complete list of tasks, while Supplementary Table 2 summarizes the categories of these tasks. All diagnostic subcategories including “Diagnostic: Tumor Staging,” “Diagnostic: Tumor Grading,” “Diagnostic: Histologic Subtyping,” “Diagnostic: Primary Diagnosis,” “Diagnostic: Tissue Type Classification,” “Diagnostic: Primary Diagnosis & Histologic Subtyping,” “Diagnostic: Primary Diagnosis & Tumor Grading,” and “Diagnostic: Other”—were consolidated under the broader category of ‘Diagnostic’. The description of 41 tasks are summarized below: TCGA LGG & GBM IDHT mutation status prediction The TCGA histopathology images were downloaded from GDC portal, while clinical and molecular information of patients with the ground truth IDH mutation status was downloaded from the TCGA glioblastoma study77. This binary slide-level classification task included classification of samples as IDH mutated vs wild type (WT), directly from an input H&E stained whole-slide image of the patient’s brain tumor. The training subset contained N = 884, while the hold-out test set comprised N = 368 samples. TCGA LGG & GBM MGMT methylation status prediction This binary slide-level classification task consisted of predicting MGMT methylation status in TCGA glioblastoma study patients, directly from input H&E stained whole-slide image of patient’s brain tumor. The ground truth MGMT methylation status was collected from the TCGA glioblastoma study. The training subset consisted of N = 480 samples. The hold-out test set contained N = 200 samples. TCGA LGG & GBM histologic subtype classification This multi-class slide-level classification task consisted of predicting the TCGA GBM & LGG histologic subtypes: astrocytoma, glioblastoma, oligoastrocytoma, oligodendroglioma, directly from input H&E stained whole-slide image of the patient’s brain tumor. There were N = 192 in training, N = 80 in hold out test sets. The ground truth histology subtype information was also collected from the TCGA glioblastoma study. TCGA LGG & GBM transcriptomic subtype prediction This multi-class slide-level classification task consisted of predicting the TCGA GBM & LGG transcriptomic subtypes: Classical (CN), Mesenchymal (ME), Neural (NE), Proneural (PN), directly from an input H&E stained whole-slide image of the patient’s brain tumor. The ground truth transcriptome subtype information was also collected from the TCGA glioblastoma study. The training and hold-out test set consisted of N = 744, N = 308 samples respectively. TCGA BRCA ER status prediction TCGA BRCA tumor clinical and molecular information was obtained from the TCGA gynecologic and breast cancer study81. This binary slide-level classification task consisted of predicting the ER-positive versus ER-negative status directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth ER status was reported to be obtained by IHC in this study. There were N = 280 in training, N = 117 test samples. TCGA BRCA PR status prediction TCGA BRCA tumor clinical and molecular information was obtained from TCGA gynecologic and breast cancer study81. This binary slide-level classification task consisted of predicting the PR-positive versus PR-negative status directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth PR status was determined using immunohistochemistry (IHC) as reported in the original study. The dataset was split into N = 413 samples for training and N = 172 for testing. TCGA BRCA PAM50 status prediction This multi-class slide-level classification task consisted of predicting the PAM50 subtype, as defined in the original study81, directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth information subtype information for the subtypes—Basal, HER2-enriched (HER2E), Luminal A, and Luminal B—were all sufficiently represented. The dataset was split into N = 194 training samples and N = 81 testing samples. TCGA BLCA histologic subtype classification In this task, we performed binary slide-level classification of H&E-stained whole-slide images from TCGA BLCA as either papillary or non-papillary tumors. The ground truth histologic subtype information was obtained from the TCGA BLCA study82. The dataset was split into N = 153 training samples and N = 64 testing samples. TCGA BLCA mRNA cluster prediction In this task, we performed multi-class slide-level classification (3-class) of predicting mRNA expression clusters in TCGA BLCA, distinguishing among Basal/Squamous, Luminal Infiltrated, and Luminal Papillary subtypes, directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth mRNA cluster information was obtained from the TCGA BLCA study82. The dataset was split into 160 training samples and 67 testing samples. TCGA gastrointestinal adenocarcinoma MSI status prediction We performed a three-class slide-level classification task to predict MSI status (MSI-high, MSI-low, MSS) using samples from TCGA gastrointestinal tumors, including TCGA-COAD, TCGA-READ, TCGA-STAD, and TCGA-ESCA, directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth MSI status of patients were obtained from the TCGA gastrointestinal tumor study85. The dataset was divided into N = 451 training samples and N = 189 testing samples. TCGA LUSC vs LUAD classification In this task, we performed binary slide-level classification of H&E-stained whole-slide images of non-small-cell lung cancers in the TCGA lung cancer study into the squamous cell carcinoma (LUSC) versus lung adenocarcinoma (LUAD) histologic subtypes. The ground truth lung cancer subtype information was obtained from the TCGA lung cancer study79,80. The dataset was split into N = 1453 training samples and N = 607 testing samples. TCGA lung transcriptomic subtype prediction This multi-class (5-class) slide-level classification task involved predicting 5 transcriptional subtypes: basal, classical, proximal inflammatory, proximal proliferative & terminal respiratory unit(TRU), in TCGA LUSC & TCGA LUAD tumors, directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth lung transcriptomic subtype information was obtained from a lung expression subtype study86. The dataset was split into N = 395 training samples and N = 165 testing samples. TCGA kidney tumor subtype classification This multi-class (4-class) slide-level classification task consisted of determining the histologic subtype of H&E-stained whole-slide images from TCGA renal cell carcinomas. The 4 classes were ccRCC (Clear cell renal cell carcinoma), ChRCC (chromophobe renal cell carcinoma), PRCC1 (Papillary renal cell carcinoma type 1), and PRCC2 (Papillary renal cell carcinoma type 2). The ground truth histologic subtype information of these tumors was obtained from the TCGA renal cell study87. The train and test split was N = 194, N = 80 samples. TCGA pan-cancer grading In this multi-class slide-level classification task, we performed three-class tumor grading (G1 vs G2 vs G3) on input H&E-stained whole-slide images from multiple tumor types. The ground truth tumor grade was obtained from the TCGA Pan-Cancer Clinical Data Resource88. The dataset was split into N = 717 training samples and N = 300 testing samples. TCGA pan-cancer PI3K pathway activation status prediction This binary slide-level classification task consisted of predicting PI3K pathway activation status (active PI3K signaling vs. no PI3K pathway activity) across TCGA pan-cancer samples, directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth PI3K pathway activation status was derived from the TCGA Oncogenic Signaling Pathways study83. The train and test was split into N = 6244, N = 2603 samples. TCGA pan-cancer MYC pathway activation status prediction This binary slide-level classification task consisted of predicting MYC pathway activation status (active MYC signaling vs. no MYC pathway activity) from H&E-stained whole-slide images from TCGA pan-cancer samples. The ground truth MYC pathway activation status was derived from the TCGA Oncogenic Signaling Pathways study83. The train and test was split into N = 2308, N = 963 samples. TCGA pan-cancer immune sub-group prediction In this slide-level task performed across TCGA pan-cancer samples, we predicted which of the following six classes a sample fell into, based on an input H&E-stained whole-slide image from the patient’s tumor: C1 (Wound Healing), C2 (IFN-γ Dominant), C3 (Inflammatory), C4 (Lymphocyte Depleted), C5 (Immunologically Quiet), and C6 (TGF-β Dominant). The ground truth classification was obtained from the integrated immunogenomics study from the TCGA immune landscape publication84. The dataset was split into N = 996 training samples, and N = 416 testing samples. TCGA PRAD ERG fusion status prediction In this binary slide-level classification task on the TCGA PRAD dataset, the ERG fusion status (positive vs negative) was predicted directly from an input H&E-stained whole-slide image of the patient’s tumor. The ground truth ERG fusion status was obtained from the TCGA PRAD study78. The dataset was split into N = 405 training samples, and N = 169 testing samples. TCGA PRAD mRNA clustering In this multi-class (3-class) slide-level classification task, we predicted the mRNA expression cluster (CL1, CL2, or CL3) directly from input H&E-stained whole-slide images from the TCGA PRAD (prostate adenocarcinoma) dataset78. According to the original study, CL1 primarily comprised ETS-negative tumors, while CL2 and CL3 represented ETS-positive tumors, with CL3 additionally enriched for PI3K/AKT pathway activity. The ground truth mRNA expression cluster information for patients was obtained from the TCGA PRAD study. The dataset was split into N = 413 for training, and N = 173 for testing. CPTAC pan-cancer grading We performed multi-class (4-class) slide-level classification of the tumor grade on input H&E-stained whole-slide images from the CPTAC pan-cancer study65. The ground truth tumor grades obtained from the study were G1, G2, G3, and G4. The train and test samples were N = 168 and N = 71, respectively. CPTAC pan-cancer tumor type classification We performed multi-class (7-class) slide-level classification of the tumor types on input H&E-stained whole-slide images from the CPTAC pan-cancer study66. The ground truth tumor type classes (BRCA, CCRCC, COAD, GBM, LSCC, LUAD, PDA) were obtained from the CPTAC pan-cancer study. The train and test samples were N = 441 and N = 175 samples respectively. CPTAC pan-cancer chromatin modification pathway activation prediction In this binary slide-level classification task, we predicted the chromatin modification pathway activation status (activated vs non-activated) from H&E-stained whole-slide images across 7 cancer types from the CPTAC pan-cancer study66. The ground truth status for chromatin modification pathway activation was obtained from the CPTAC pan-cancer study. The train and test samples were N = 472 and N = 198 samples respectively. CPTAC pan-cancer mTOR pathway activation prediction In this binary classification task, we predicted the mTOR pathway activation status (activated vs non-activated) from H&E-stained whole-slide images across 7 cancer types from the CPTAC pan-cancer study66. The ground truth status for mTOR pathway activation was obtained from the CPTAC pan-cancer study. The train and test samples were N = 532 and N = 222, respectively. CPTAC pan-cancer MYC pathway activation prediction In this binary slide-level classification task, we predicted the MYC pathway activation status (activated vs non-activated) from H&E-stained whole-slide images across 7 cancer types from the CPTAC pan-cancer study66. The ground truth status for MYC pathway activation was obtained from the CPTAC pan-cancer study. The train and test samples were N = 860 and N = 358, respectively. CPTAC pan-cancer p53 pathway activity In this binary slide-level classification task, we predicted the p53 pathway activation status (activated vs non-activated) from H&E-stained whole-slide images across 7 cancer types from the CPTAC pan-cancer study66. The ground truth status for p53 pathway activation was obtained from the CPTAC pan-cancer study. The train and test samples were N = 488 and N = 204, respectively. CPTAC pan-cancer SWI-SNF pathway activity In this binary slide-level classification task, we predicted the SWI/SNF (Switch/Sucrose Non-fermenting) pathway activation status (activated vs non-activated) from H&E-stained whole-slide images across 7 cancer types from the CPTAC pan-cancer study66. The ground truth status for SWI/SNF pathway activation was obtained from the CPTAC pan-cancer study. The train and test samples were N = 524 and N = 218, respectively. NLST (national lung screening trial) tumor stage classification (out of domain) In this multi-class slide-level classification task, we predicted lung cancer clinical stage (stage I, II, or III) using H&E whole slide images from the TCIA National Lung Screening Trial (NLST, https://www.cancerimagingarchive.net/collection/nlst/)) study67. The stages were defined according to AJCC staging criteria in the original study67. The ground truth clinical stage labels were obtained from the NLST study. The dataset was split into training (N = 76) and test (N = 52) sets. NLST (national lung screening trial) tumor grade classification (out of domain) In this multi-class slide-level classification task, we predicted lung cancer histological grades(Grade 1, Grade 2 or Grade 3) using H&E whole slide images from the TCIA National Lung Screening Trial (NLST) study67. The three tumor grades classified were Grade 1 (well-differentiated), Grade 2 (moderately differentiated), and Grade 3 (poorly differentiated), based on standard histopathological criteria. The ground truth information on tumor grades was obtained from the NLST study. The dataset (https://www.cancerimagingarchive.net/collection/nlst/) was divided into training (N = 138) and test (N = 94) sets. DHMC lung adenocarcinoma classification (out of domain) In this multi-class slide-level classification task, we predicted lung adenocarcinoma histological patterns using H&E whole slide images from the Dartmouth-Hitchcock Medical Center (DHMC) dataset (https://bmirds.github.io/LungCancer/)68. The three predominant histological pattern classes were lepidic, acinar, and solid, based on standard histopathological criteria for lung adenocarcinoma grading. This classification is critical for determining tumor grade and guiding treatment decisions, though it is challenging due to the heterogeneous nature of lung adenocarcinoma and subjective evaluation criteria. The ground truth information on DHMC lung adenocarcinoma classification was directly obtained from the DHMC dataset68. The dataset was divided into training (N = 66) and test (N = 46) sets. Stanford GBM MGMT methylation status (out of domain) This binary slide-level classification task consisted of predicting MGMT promoter methylation status in glioblastoma (GBM) using H&E-stained whole-slide images from the Stanford GBM dataset. The two classes distinguished samples with methylated versus unmethylated MGMT promoters. The ground truth MGMT methylation status was obtained from the Stanford clinical database. The dataset was divided into training (N = 45) and test (N = 31) sets. Stanford GBM necrosis prediction (out of domain) This binary slide-level classification task consisted of predicting tumor necrosis status in glioblastoma (GBM) using H&E-stained whole-slide images from the Stanford GBM dataset. The two classes distinguished samples with necrosis versus those without necrosis. The ground truth necrosis status was obtained from the Stanford clinical database. The dataset was divided into training (N = 49) and test (N = 34) sets. Stanford lung tumor histology classification (out of domain) In this slide-level task, we performed binary classification of H&E-stained whole-slide images of non-small-cell lung cancers in the Stanford lung cancer dataset into the lung adenocarcinoma (LUAD) versus squamous cell carcinoma (LUSC) histologic subtypes. The ground truth histology status was obtained from the Stanford clinical database. The dataset was split into N = 29 training samples and N = 21 testing samples. Stanford lung tumor stage classification (out of domain) In this slide-level task, we performed multi-class classification of H&E-stained whole-slide images of lung cancers in the Stanford lung cancer dataset into clinical stages I, II, and III. The ground truth stage information was obtained from the Stanford clinical database. The dataset was split into N = 28 training samples and N = 19 testing samples. BACH (external benchmarking)70 In this patch-level multi-class classification task, we predicted breast histology patterns using data from the Breast Histology Images (BACH) dataset (https://zenodo.org/records/3632035), from hematoxylin and eosin (H&E) stained microscopy images. The four classes were normal, benign, in situ carcinoma, and invasive carcinoma, representing key diagnostic categories for determining breast cancer type and guiding treatment decisions. The ground truth information of breast histology classes was obtained from the BACH website70. The dataset was divided into training (N = 280 images) and validation (N = 120 images) sets. BRACS (external benchmarking)69 In this patch-level multi-class classification task, we predicted breast lesion subtypes from the BRACS dataset (https://www.bracs.icar.cnr.it/) from hematoxylin and eosin (H&E) stained histological images. The seven classes classified included normal tissue (N), pathological benign (PB), usual ductal hyperplasia (UDH), flat epithelial atypia (FEA), atypical ductal hyperplasia (ADH), ductal carcinoma in situ (DCIS), and invasive carcinoma (IC). The ground truth information of the 7 classes was obtained from BRACS website69. The dataset was divided into training (N = 3670 images) and validation (N = 325 images) sets. BreakHis (external benchmarking)73 We conducted patch-based binary classification to differentiate benign from malignant breast tumor tissue using data from the BreakHis dataset (https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/). The classification identified histopathological patterns in hematoxylin and eosin (H&E) stained microscopic images, captured at multiple magnifications (40X, 100X, 200X, 400X). This task is essential for precise breast cancer diagnosis and treatment planning, yet challenging due to histopathological variability across magnifications. Developed in collaboration with the P&D Laboratory – Pathological Anatomy and Cytopathology, Paraná, Brazil, to advance automated breast cancer histopathology research, the dataset was split into training (N = 5008 images) and validation (N = 2907 images) sets.The ground truth information was obtained from BreakHis website73. Lung and colon cancer histopathological (LC25000)74 (external benchmarking) In this multi-class patch based classification task, we predicted lung and colon histopathological patterns using H&E whole slide image data from the LC25000 dataset (https://github.com/tampapath/lung_colon_image_set). The five classes classified were lung adenocarcinoma, lung squamous cell carcinoma, benign lung tissue, colon adenocarcinoma, and benign colon tissue, based on hematoxylin and eosin (H&E) stained microscopic images. This classification is critical for accurate diagnosis and treatment planning in lung and colon cancers, though challenging due to histopathological similarities across malignant and benign tissues. The ground truth information of the 5 classes was obtained from the LC25000 website. The dataset was divided into training (N = 17,500 images) and validation (N = 7500 images) sets74. NCT-CRC-HE (external benchmarking) In this multi-class patch based classification task, we predicted colorectal tissue histopathological patterns using data from the NCT-CRC-HE-100K dataset (https://zenodo.org/records/1214456). The nine classes classified were adipose (ADI), background (BACK), debris (DEB), lymphocytes (LYM), mucus (MUC), smooth muscle (MUS), normal colon mucosa (NORM), cancer-associated stroma (STR), and colorectal adenocarcinoma epithelium (TUM), based on hematoxylin and eosin (H&E) stained image patches (224×224 pixels, 0.5 microns per pixel). This classification is critical for accurate colorectal cancer diagnosis and understanding tumor microenvironments, though challenging due to histopathological variability across tissue types. The ground truth information was obtained from the NCT-CRC-HE dataset76. The dataset, sourced from the NCT Biobank and UMM pathology archive, was divided into training (N = 100,000 images), validation (N = 7180 images). MHIST75(External benchmarking) We executed binary patch based classification to distinguish colorectal polyp histopathological patterns using data from the MHIST dataset. The two classes identified were hyperplastic polyp (HP) and sessile serrated adenoma (SSA), based on hematoxylin and eosin (H&E) stained microscopic images. This classification is vital for differentiating benign HPs from precancerous SSAs, which require earlier clinical follow-up due to their potential to progress to colorectal cancer. The ground truth information for two classes was obtained from MHIST dataset75. The dataset (https://bmirds.github.io/MHIST/) was partitioned into training (N = 2178 images) and validation (N = 980 images) sets. SICAPv2 Gleason grading89(External benchmarking) We executed multi-class patch-level classification (four classes) of predicting prostate cancer histopathological patterns using data from the SICAPv2 dataset. The four classes identified were non-cancerous (NC), Gleason grade 3 (GG3), Gleason grade 4 (GG4), and Gleason grade 5 (GG5), based on hematoxylin and eosin (H&E) stained whole-slide images digitized at 40x magnification. This classification is vital for assessing prostate cancer aggressiveness and guiding treatment decisions. The ground truth information was obtained from SICAPv2 dataset89. The dataset (https://data.mendeley.com/datasets/9xxm58dvs3/1), sourced from the Hospital Clínico of Valencia and designed to advance research in automated prostate cancer histopathology, was partitioned into training (N = 9962 images) and validation (N = 2125 images) sets. UniToPatho71(external benchmarking) We executed multi-class patch-level classification (six classes) to predict colorectal histopathological patterns using H&E stained whole slide image data from the UniToPatho dataset (https://ieee-dataport.org/open-access/unitopatho). The six classes identified were normal tissue (NORM), hyperplastic polyp (HP), tubular adenoma with high-grade dysplasia (TA.HG), tubular adenoma with low-grade dysplasia (TA.LG), tubulo-villous adenoma with high-grade dysplasia (TVA.HG), and tubulo-villous adenoma with low-grade dysplasia (TVA.LG), based on hematoxylin and eosin (H&E) stained image patches digitized at 20× magnification (0.4415 μm/pixel). This classification is essential for distinguishing benign from precancerous colorectal lesions and grading adenomas to guide clinical management. The ground truth information was obtained from UniToPatho dataset90. The dataset, sourced from 292 whole-slide images and curated to advance research in automated colorectal histopathology, was partitioned into training (N = 6281 images) and validation (N = 2410 images) sets. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 -15 Reporting Summary Transparent Peer Review file