A Comprehensive Benchmark of Histopathology Foundation Models for Kidney Digital Pathology Images

Harishwar Reddy Kasireddy; Patricio S. La Rosa; Akshita Gupta; Anindya S. Paul; Jamie L. Fermin; William L. Clapp; Meryl A. Waldman; Tarek M. El-Ashkar

cs.CVMar 16, 2026

A Comprehensive Benchmark of Histopathology Foundation Models for Kidney Digital Pathology Images

Authors: Harishwar Reddy Kasireddy, Patricio S. La Rosa, Akshita Gupta, Anindya S. Paul, Jamie L. Fermin, William L. Clapp, Meryl A. Waldman, Tarek M. El-Ashkar, +7 more

Organizations: Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL, USA · Seed Production Innovation, Crop Science Division, Bayer Company, St. Louis, MO, USA · Division of Medicine – Quantitative Health, University of Florida, Gainesville, FL, USA · Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL, USA · Department of Pathology, Immunology and Laboratory Medicine, University of Florida College of Medicine, Gainesville, FL, USA · Kidney Disease Branch, National Institute of Diabetes and Digestive and Kidney Diseases, National Institutes of Health, Bethesda, MD, USA · Indiana University School of Medicine, Indianapolis, IN, USA · Departments of Medicine, Washington University School of Medicine, St. Louis, MO, USA · Universidade de Coimbra, Coimbra, Coimbra District, Portugal · Department of Pathology and Laboratory Medicine, University of California at Davis School of Medicine, Sacramento, CA, USA · Department of Pathology, Johns Hopkins University School of Medicine, Baltimore, MD, USA · Department of Pathology, University of Michigan, Ann Arbor, MI, USA

Abstract

Histopathology foundation models (HFMs), pretrained on large-scale cancer datasets, have advanced computational pathology. However, their applicability to non-cancerous chronic kidney disease remains underexplored, despite coexistence of renal pathology with malignancies such as renal cell and urothelial carcinoma. We systematically evaluate 11 publicly available HFMs across 11 kidney-specific downstream tasks spanning multiple stains (PAS, H&E, PASM, and IHC), spatial scales (tile and slide-level), task types (classification, regression, and copy detection), and clinical objectives, including detection, diagnosis, and prognosis. Tile-level performance is assessed using repeated stratified group cross-validation, while slide-level tasks are evaluated using repeated nested stratified cross-validation. Statistical significance is examined using Friedman test followed by pairwise Wilcoxon signed-rank testing with Holm-Bonferroni correction and compact letter display visualization. To promote reproducibility, we release an open-source Python package, kidney-hfm-eval, available at https://pypi.org/project/kidney-hfm-eval/ , that reproduces the evaluation pipelines. Results show moderate to strong performance on tasks driven by coarse meso-scale renal morphology, including diagnostic classification and detection of prominent structural alterations. In contrast, performance consistently declines for tasks requiring fine-grained microstructural discrimination, complex biological phenotypes, or slide-level prognostic inference, largely independent of stain type. Overall, current HFMs appear to encode predominantly static meso-scale representations and may have limited capacity to capture subtle renal pathology or prognosis-related signals. Our results highlight the need for kidney-specific, multi-stain, and multimodal foundation models to support clinically reliable decision-making in nephrology.

Figures & tables

Model Name	Encoder Type	Dataset Size	Pretraining Image Type	SSL Strategy	Image Magnification	Number of Parameters	Feature Dimension
H-optimus-0 [ 95 ]	ViT-G	500k $\dagger$	H&E	DINOv2 [ 83 ]	20x	1.10B	1536*
H-optimus-1 [ 8 ]	ViT-G	1M $\dagger$	H&E	DINOv2	20x	1.10B	1536*
Hibou-B [ 79 ]	ViT-B	512M $\dagger\dagger$	H&E, non-H&E, and cytology	DINOv2	Not reported	86M	768*
Hibou-L [ 79 ]	ViT-L	1.2B $\dagger\dagger$	H&E, non-H&E, and cytology	DINOv2	Not reported	303M	1024*
Prov-Gigapath [ 115 ]	ViT-G + LongNet [ 28 ]	171k $\dagger$	H&E and IHC	DINOv2 + MAE [ 37 ]	20x	1.10B	1536, 768*
SP22M [ 13 ]	ViT-S	423k $\dagger$	H&E	DINO [ 15 ]	20x	22M	384**

Table 1: Summary of vision histopathology foundation models (HFMs) evaluated in this study. Each model varies in architecture type, pretraining dataset scale (no. of tiles), pretraining image modality (i.e., type of stain), SSL strategy, magnification levels of images and their resolution (i.e., number of pixels), number of HFM parameters, and output feature dimensionality. Note: * indicates tile-level embedding dimension; ** indicates slide-level embedding dimension;

\dagger

indicates number of WSIs;

\dagger\dagger

indicates number of tiles.

Evaluation Task	Level	Task type	Number of Slides	Number of tiles	Data source	Number of classes	Class ratio	Staining Protocol
Globally vs. non-globally sclerotic glomerulus classification	Tile	Classification	215	5761	WUSTL, Vanderbilt University	2	1:4.46	PAS
GBM vs. no GBM spike classification	Tile	Classification	63	695	NIH/ NIDDK	2	1:1.93	PASM
Reference vs. abnormal tubule classification	Tile	Classification	12	1200	Indiana University	2	1:1	H&E
Inflammation vs. no inflammation classification	Tile	Classification	14	8245	UC Davis	2	1.26:1	H&E
Arterial stenosis classification (mild, moderate, severe)	Tile	Classification	155	242	University of Coimbra, UC Davis	3	5.32:2.08:1	PAS
Cell type proportion estimation	Tile	Regression	12	17,478	Indiana University	NA	NA	H&E

Table 2: Summary of datasets being used for evaluating histopathology foundation models. Abbreviations: WUSTL = Washington University in St. Louis, NIH = National Institutes of Health, NIDDK = National Institute of Diabetes and Digestive and Kidney Diseases, UC Davis = University of California Davis, JHU = Johns Hopkins University, NA = Not Applicable, GBM = Glomerular Basement Membrane, eGFR = estimated Glomerular Filtration Rate, MN = Membranous Nephropathy, DN = Diabetic Nephropathy.

Panel of Figure 1 — Figure 1: Representative histopathology examples of glomerulosclerosis.

Panel of Figure 2 — Figure 2: Representative histopathology examples of GBM spike detection.

Panel of Figure 3 — Figure 3: Representative histopathology examples reference and abnormal tubules.

Panel of Figure 4 — Figure 4: Representative histopathology examples with and without inflammation.

Panel of Figure 5 — Figure 5: Representative histopathology examples of mild, moderate, and severe arteriolar stenosis.

Panel of Figure 6 — Figure 6: Sample histology image patch from which the cell type is estimated.

Panel of Figure 7 — Figure 7: Representative histopathology examples of the augmentations applied to a single microvascular inflammation tissue patch for the copy detection task.

Figure 8: Overview of the tile-level evaluation workflow for benchmarking histopathology foundation models (HFMs). (A) An input histology tile is passed through one of 11 HFM encoders (H-optimus-0, H-optimus-1, Hibou-B, Hibou-L, Prov-Gigapath, SP22M, SP85M, UNI, UNI2-h, Virchow, Virchow2) to obtain a fixed-dimensional feature embedding. (B) The extracted embedding is used to train lightweight machine learning models either linear or $k$ NN probing for classification tasks, or ridge regression for continuous prediction tasks. (C) Model performance is evaluated using a repeated group five-fold cross-validation framework, followed by bootstrapping of predictions, and a statistical framework integrating Friedman test, Wilcoxon signed-rank tests, and Compact Letter Display (CLD) [ 35 ] analysis to quantify variance and compute generalized predictive performance across all tile-level tasks.

Figure 9: Overview of the slide-level evaluation framework for benchmarking histopathology foundation models (HFMs). (A) Whole-slide images (WSIs) are tiled and filtered to generate high-quality image patches. (B) Each tile is passed through one of 11 HFM encoders (H-optimus-0, H-optimus-1, Hibou-B, Hibou-L, Prov-Gigapath, SP22M, SP85M, UNI, UNI2-h, Virchow, Virchow2) to obtain fixed-dimensional feature embeddings. (C) Tile embeddings belonging to the same slide are aggregated using an attention-based multiple instance learning (ABMIL) model, which is trained for either regression or classification tasks. (D) ABMIL hyperparameter tuning is conducted using a repeated nested stratified five-fold cross-validation framework, where inner folds are used for hyperparameter tuning and outer folds provide unbiased test estimates. (E) Predictions from all folds are bootstrapped, followed by a statistical framework integrating the Friedman test, Wilcoxon signed-rank tests, and Compact Letter Display (CLD) analysis to quantify variance and compute generalized predictive performance across slide-level tasks.

Table 3: Performance comparison of HFMs for global glomerulosclerosis classification. Tables 3(a) and 3(b) rank the performance of 11 HFMs for global glomerulosclerosis classification using logistic regression and

k

-nearest neighbors classifiers, respectively. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated stratified group five-fold cross-validation runs. Statistical ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Table 4: Performance comparison of HFMs for glomerular basement membrane spike classification. Tables 4(a) and 4(b) rank the performance of 11 HFMs for glomerular basement membrane spike classification using logistic regression and

k

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Table 5: Performance comparison of HFMs for tubular classification. Tables 5(a) and 5(b) rank the performance of 11 HFMs for tubular classification using logistic regression and

k

-nearest neighbors classifiers, respectively. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated stratified group five-fold cross-validation runs. Statistical Ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Table 6: Performance comparison of HFMs for inflammation classification. Tables 6(a) and 6(b) rank the performance of 11 HFMs for inflammation classification using logistic regression and

k

-nearest neighbors classifiers, respectively. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated stratified group five-fold cross-validation runs. Statistical Ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Table 7: Performance comparison of HFMs for arterial stenosis classification. Tables 7(a) and 7(b) rank the performance of 11 HFMs for arterial stenosis classification using logistic regression and

k

-nearest neighbors classifiers, respectively. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated stratified group five-fold cross-validation runs. Statistical Ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

CLD	Model	Min	1st Quart.	Median	Mean	3rd Quart.	Max
a	UNI2-h	0.3533	0.3612	0.3634	0.3635	0.3657	0.3737
b	H-optimus-1	0.3286	0.3364	0.3388	0.3387	0.3411	0.3487
c	H-optimus-0	0.3035	0.3129	0.3150	0.3150	0.3171	0.3245
d	UNI	0.2910	0.3017	0.3042	0.3041	0.3063	0.3139
e	Virchow2	0.2797	0.2886	0.2910	0.2910	0.2934	0.3025
f	Virchow	0.2760	0.2836	0.2858	0.2858	0.2881	0.2986

Table 8: Performance comparison of HFMs for cell type estimation. Table 8 ranks the performance of 11 HFMs for cell type estimation using ridge regression. Performance is reported as the Pearson Correlation Coefficient (PCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated group five-fold cross-validation runs. Ranking of model performance for cell type estimation. Statistical Ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Augmentation	H-optimus-0	H-optimus-1	Hibou-B	Hibou-L	Prov-Gigapath	SP22M	SP85M	UNI	UNI2-h	Virchow	Virchow2
color	0.9343	0.9359	0.8997	0.8735	0.9538	0.8669	0.8798	0.9480	0.9397	0.9255	0.9555
deform	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	0.9717	0.9995
geo	0.9979	0.9954	0.9891	0.8827	1.0000	0.9987	0.9991	0.9987	0.9879	0.9663	0.9575
noise	0.8124	0.6715	0.2457	0.4324	0.8656	0.8074	0.6307	0.8112	0.7758	0.6216	0.6199

Table 9: Top-1 classification accuracy for HFMs under four augmentation types (color, deform, geometric, noise), evaluating representation robustness.

CLD	Model	Min	1st Quart.	Median	Mean	3rd Quart.	Max
a	H-optimus-0	0.7266	0.9270	1.0000	0.9685	1.0000	1.0000
b	Virchow	0.6574	0.9220	1.0000	0.9538	1.0000	1.0000
c	Virchow2	0.5883	0.9117	0.9282	0.9381	1.0000	1.0000
d	SP85M	0.5416	0.9035	0.9250	0.9269	1.0000	1.0000
d	UNI2-h	0.6064	0.9035	0.9250	0.9268	1.0000	1.0000
d	UNI	0.6064	0.8748	0.9250	0.9240	1.0000	1.0000

Table 10: Performance comparison of HFMs for healthy reference samples vs. DN classification. Table 10 ranks the performance of 11 HFMs for healthy reference samples vs. DN classification using attention based multiple instance learning (ABMIL) classifier. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated nested stratified five-fold cross-validation runs. Statistical ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

CLD	Model	Min	1st Quart.	Median	Mean	3rd Quart.	Max
a	H-optimus-1	-0.2025	0.0000	0.0000	0.0488	0.0000	1.0000
a	UNI2-h	-0.2988	-0.0913	0.0000	0.0484	0.0000	1.0000
b	Virchow2	-0.2988	-0.0913	0.0000	0.0148	0.0000	1.0000
b	Hibou-L	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
b	Hibou-B	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
c	UNI	-0.4767	-0.1333	-0.0625	-0.0099	0.0000	1.0000

Table 11: Performance comparison of HFMs for MN treatment response classification. Table 11 ranks the performance of 11 HFMs for MN treatment response classification using attention based multiple instance learning (ABMIL) classifier. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated nested stratified five-fold cross-validation runs. Statistical Ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

CLD	Model	Min	1st Quart.	Median	Mean	3rd Quart.	Max
a	Hibou-B	-0.5449	-0.0339	0.1038	0.1039	0.2381	0.7542
b	Virchow2	-0.5449	-0.1340	0.0000	0.0169	0.1581	0.6325
b	H-optimus-1	-0.6172	-0.1372	0.0000	0.0012	0.1372	0.7688
b	Virchow	-0.7006	-0.1494	-0.0109	-0.0060	0.1340	0.6720
c	Hibou-L	-0.5709	-0.1682	-0.0369	-0.0393	0.0895	0.5916
d	UNI	-0.6231	-0.2186	-0.0714	-0.0687	0.0778	0.6706

Table 12: Performance comparison of HFMs for eGFR decline classification in post-renal transplantation. Table 12 ranks the performance of 11 HFMs for eGFR decline classification in post-renal transplant cases using an attention-based multiple instance learning (ABMIL) classifier. Performance is reported as the Matthews Correlation Coefficient (MCC) statistics (minimum to maximum) across 1000 bootstraps obtained from repeated nested stratified five-fold cross-validation runs. Statistical ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

CLD	Model	Min	1st Quart.	Median	Mean	3rd Quart.	Max
a	Hibou-L	-0.4802	-0.1604	-0.0808	-0.0925	-0.0133	0.2477
b	Hibou-B	-0.6246	-0.2143	-0.1234	-0.1320	-0.0372	0.2657
c	Virchow2	-1.0401	-0.3271	-0.1756	-0.1888	-0.0375	0.3247
d	UNI2-h	-0.9761	-0.3342	-0.1883	-0.2131	-0.0713	0.2603
e	SP22M	-1.0089	-0.3750	-0.2259	-0.2453	-0.0991	0.2664
f	SP85M	-0.9875	-0.4216	-0.2682	-0.2827	-0.1330	0.3867

Table 13: Performance comparison of HFMs for one-year eGFR prediction in post-renal transplantation. Table 13 ranks the performance of 11 HFMs for eGFR prediction using an attention-based multiple instance learning (ABMIL) regressor. Performance is reported as the

R^{2}

statistics (minimum to maximum) across 1000 bootstraps obtained from repeated nested five-fold cross-validation runs. Statistical ranking using Compact Letter Display (CLD): The ’CLD’ column groups models based on their statistical similarity. Models sharing a letter (e.g., ’a’) are not significantly different from one another (

p>.05

). Models with different letters (e.g., ’a’ vs. ’b’) have significantly different performance (

p<.05

, Wilcoxon signed rank test).

Panel of Figure 10 — Figure 10: Violin plots showing the distribution of Matthews correlation coefficient (MCC) values across HFMs for multiple downstream tasks. Each subplot represents a specific task and probing method: (a)–(b) globally vs. non-globally sclerotic glomeruli classification using logistic regression (LR) and $k$ NN, (c)–(d) GBM spike vs. no-GBM spike classification (LR, $k$ NN), (e)–(f) normal vs. abnormal tubule classification (LR, $k$ NN), (g)–(h) inflammation vs. non-inflammation classification (LR, $k$ NN), (i)–(j) arteriolar stenosis multi-class classification (LR, $k$ NN), (k) cell type estimation (ridge regression), (l) diabetic nephropathy vs. control classification using MIL. (m) membranous nephropathy treatment-response prediction using multiple-instance learning (MIL), (n) renal transplant eGFR decline classification using MIL, (o) renal transplant eGFR prediction using MIL. Wider violins indicate greater performance variability, and higher median lines indicate more consistent model performance across cross-validation replicates.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Supplementary Figure 1 : Pipeline for generating ground truth labels for tubules using spatial transcriptomics and histology. (A) H&E-stained histology image patch with segmented tubules ( $T_{i}$ ) and Visium spatial transcriptomics spot locations ( $S_{j}$ ). (B) Seurat-based cell deconvolution workflow: kidney reference single-cell RNA-seq data (cell-by-gene matrix with cell-type annotations) is integrated with Visium spot-level gene expression (spot-by-gene matrix) to estimate cell-type proportions per spatial spot. (C) Deconvolved Visium spots overlaid on the histology image, each showing estimated proportions of multiple cell types. (D) Assignment of each spatial spot to its nearest tubule segment transferring spot-specific cell-type proportions to the corresponding tubule. (E) Final tubule-level ground truth labels, obtained by taking weighted ( $w_{i}$ ) combination of $y_{i}$ where each tubule ( $T_{i}$ ) is associated with its aggregated cell-type proportion vector ( $\mathbf{y}_{T_{k}}$ ) for downstream task.

Category	Hyperparameter	Value(s)
Model / Architecture	Instance embedding dimension ( $M$ )	$\{256,\;512,\;1024\}$
	Attention hidden dimension ( $L$ )	$\{32,\;128,\;256\}$
	Dropout	$0.6$ (fixed)
Optimization	Optimizer	AdamW
	Learning rate ( $\eta$ )	$\{5e-5,\;1e-4,\;2e-4\}$
	Weight decay	$1e-2$ (fixed)

Appendix

Supplementary Table 1: Hyperparameters used for repeated nested stratified ABMIL training for slide-level evaluation.

Panel of Figure 26 — Supplementary Figure 2 : Heatmap plots showing the p -values obtained from comparing the Matthews correlation coefficient (MCC) values across HFMs for multiple downstream tasks. Each subplot represents a specific task and probing method: (a)–(b) globally vs. non-globally sclerotic glomeruli classification using logistic regression (LR) and $k$ NN, (c)–(d) GBM spike vs. no-GBM spike classification (LR, $k$ NN), (e)–(f) normal vs. abnormal tubule classification (LR, $k$ NN), (g)–(h) inflammation vs. non-inflammation classification (LR, $k$ NN), (i)–(j) arteriolar stenosis multi-class classification (LR, $k$ NN), (k) cell type estimation (ridge regression) (i) diabetic nephropathy vs. control classification using MIL. (m) membranous nephropathy treatment-response prediction using multiple-instance learning (MIL), (n) renal transplant eGFR decline classification using MIL, (o) renal transplant eGFR prediction using MIL, and Wider violins indicate greater performance variability, and higher median lines indicate more consistent model performance across cross-validation replicates.

Explore similar work

May 8, 2026

Benchmarking Foundation Models for Renal Lesion Stratification in CT

The rapid proliferation of open-source medical foundation models (FMs) raises a practical question: how well do their pre-trained representations transfer to clinically relevant but data-scarce classification tasks? Particularly in CT-based renal lesion classification, a push toward greater generalizability would be meaningful, as the field is constrained by inherently limited training data. We addressed this through a benchmark of three medical FMs on this specific task. This six-class problem spans common entities like cysts and clear cell renal cell carcinoma, alongside rare subtypes. Using a frozen feature-probing protocol, we compared FM embeddings against a handcrafted radiomics classifier and a 3D ResNet-50 trained from scratch. Models were trained on a composite dataset of 2,854 lesions and evaluated on an external test set of 234 lesions from The Cancer Imaging Archive. Our results reveal two key findings. First, FM performance (AUC 0.70-0.77) matched the from-scratch ResNet (AUC 0.72) while drastically reducing hardware demand, requiring only seconds on a CPU after feature extraction. However, the conventional radiomics baseline significantly outperformed all deep learning approaches, achieving an AUC of 0.88 (all p

\leq

0.002). This suggests that current generalist FM embeddings do not yet capture the fine-grained texture and shape heterogeneity driving histological subtype discrimination. Despite their potential in data-scarce settings, medical FMs did not surpass established models for renal lesion stratification, leaving radiomics as the current state-of-the-art.

Hartmut Häntze, Sarah de Boer, Myrthe Buser +7

Charité - Universitätsmedizin Berlin, Department of Radiology, Berlin, Germany · Radboudumc, Diagnostic Image Analysis Group, Nijmegen, The Netherlands · Klinikum rechts der Isar, TUM University Hospital, Technical University of Munich, Munich, Germany +2

Medical Imaging Foundation Models Medical Image Classification Radiomics Medical Image Representation Learning Medical Image Benchmarks

Ask PDF

Jun 9, 2026

From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology

Foundation Models (FMs) have recently redefined the state-of-the-art in histopathology by providing robust representations for whole-slide image (WSI) analysis. However, selecting the optimal foundation model (FM) for a specific clinical cohort currently requires multiple preprocessing steps, followed by computationally expensive feature extraction and the training of a Multiple Instance Learning (MIL) aggregator for every model. In this work, we investigate whether efficient tile-level linear probing can serve as a reliable proxy for slide-level performance, reducing the need to run full slide-level pipelines for every candidate encoder. We benchmark 19 state-of-the-art FMs on 42 slide-level and 16 tile-level tasks, comparing tile probing metrics against slide-level outcomes using ABMIL and Mean Pooling aggregations. We observe a high correlation between tile and slide performance across varying task difficulties, indicating that encoder representation quality is the primary determinant of WSI success. Sensitivity analyses show that transferability is stable across models and is more influenced by cohort sizes and numbers of tiles per slide than by average task difficulty. We also measure the agreement in best performing models between tile and slide-level tasks, showing tile benchmarks reliably shortlist strong candidates. Overall, our study indicates that tile-level benchmarking provides an efficient and practical first step for narrowing down candidate models, while slide-level evaluation remains essential for final validation on clinical tasks.

Sofiène Boutaj, Leo Fillioux, Maria Vakalopoulou +2

Université Paris-Saclay, CentraleSupélec, Gustave Roussy, INSERM, IHU PRISM, Cancer Data Science Unit, France · Université Paris-Saclay, CentraleSupélec, MICS Laboratory, France

Pathology Foundation Models Medical Image Benchmarks Whole-Slide Image Analysis

Ask PDF

Feb 21, 2026

Benchmarking Computational Pathology Foundation Models For Semantic Segmentation

In recent years, foundation models such as CLIP, DINO,and CONCH have demonstrated remarkable domain generalization and unsupervised feature extraction capabilities across diverse imaging tasks. However, systematic and independent evaluations of these models for pixel-level semantic segmentation in histopathology remain scarce. In this study, we propose a robust benchmarking approach to asses 10 foundational models on four histopathological datasets covering both morphological tissue-region and cellular/nuclear segmentation tasks. Our method leverages attention maps of foundation models as pixel-wise features, which are then classified using a machine learning algorithm, XGBoost, enabling fast, interpretable, and model-agnostic evaluation without finetuning. We show that the vision language foundation model, CONCH performed the best across datasets when compared to vision-only foundation models, with PathDino as close second. Further analysis shows that models trained on distinct histopathology cohorts capture complementary morphological representations, and concatenating their features yields superior segmentation performance. Concatenating features from CONCH, PathDino and CellViT outperformed individual models across all the datasets by 7.95% (averaged across the datasets), suggesting that ensembles of foundation models can better generalize to diverse histopathological segmentation tasks.

Lavish Ramchandani, Aashay Tinaikar, Dev Kumar Das +2

Aira Matrix Private Limited, Mumbai, India

Histopathology Image Segmentation Pathology Foundation Models Computational Pathology

Ask PDF