henhaobio
Sign in

The Journal of Pathology Clinical ResearchPathologyPeer-reviewed journal article

Impact of tissue staining and scanner variation on the performance of pathology foundation models: a study of sarcomas and their mimics

Binghao ChaiiD, Jianan CheniD, Paul CooliD, Fatine OumliliD, Anna Tollit, David F. SteineriD, Tapabrata ChakrabortiiD, Adrienne Margaret FlanaganiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Histopathological analysis is considered the gold standard for the diagnosis and prognostication of cancer. Recent advances in AI, driven by large-scale digitisation and pan-cancer foundation models, are opening new opportunities for clinical integration. However, it remains unclear how robust these foundation models are to real-world sources of variability, particularly in H&E staining and scanning protocols. In this study, we use soft tissue tumours, a rare and morphologically diverse tumour type, as a challenging test case to systematically investigate the colour-related robustness and generalisability of seven AI models. Controlled staining and scanning experiments were utilised to assess model performance across diverse real-world data sources. Foundation models, particularly UNI-v2, Virchow and TITAN, demonstrated encouraging robustness to staining and scanning variation, particularly when a small number of stain-varied slides were included in the training loop, highlighting their potential as adaptable and data-efficient tools for real-world digital pathology workflows.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (28 Aug 2025) · CC BY 4.0

Across 14 sarcoma and mimic subtypes, six pathology foundation models beat ResNet50 but lost accuracy under multi-institutional staining; adding 121 externally stained slides narrowed the gap.

Clinical pathology AI must perform consistently across H&E staining protocols and slide scanners, yet stain normalisation can distort morphology, GAN-based correction needs large data and transfers poorly, and foundation-model robustness to these factors is underexplored. Using rare, morphologically diverse soft tissue tumours, the study quantifies that robustness via controlled stain and scanner experiments and releases a multi-lab, multi-scanner slide collection.

Key results

  1. ShownAll six foundation models outperformed a supervised ResNet50 on slides from one lab and one scanner. Every model, however, lost significant accuracy on a real-world test set stained in many hospitals, and ResNet50 degraded most.

    Class-wise accuracy, Cohort 1-Test vs Cohort 4-Test: TITAN 0.932 vs 0.869, UNI-v2 0.939 vs 0.842, Virchow 0.921 vs 0.859, CONCH-v1.5 0.911 vs 0.737, PRISM 0.882 vs 0.811, ResNet50 0.821 vs 0.466; chi-squared decline significant for all seven models (TITAN p = 0.004, Virchow p = 0.006, others p < 0.001); best models dropped 6–10% (Discussion). · 2.1 Evaluation of foundation models for…; 3 Discussion; Figure legends

  2. ShownAdding a small set of externally stained slides to the single-lab training data significantly improved real-world accuracy for several patch-level models. It also lifted the weakest classes (nodular fasciitis, intramuscular myxoma), with a minor decline on the in-house test set for some models.

    Adding 121 Cohort 4-Train WSIs to Cohort 1-Train (Combined cohort) significantly raised Cohort 4-Test accuracy for Path Foundation (p < 0.001), UNI-v2 (p < 0.002), CONCH-v1.5 (p < 0.001) and ResNet50 (p < 0.001); not significant for Virchow (p = 0.478), TITAN (p = 0.238) or PRISM (p = 0.329); chi-squared tests. · 2.1 Evaluation of foundation models for…; 4.5 Statistical analysis; Figure legends

  3. SuggestedWith staining held constant, CONCH-v1.5, Virchow, PRISM and ResNet50 lost accuracy when trained on Hamamatsu or 3D-Histech scans and tested on Aperio scans. TITAN and UNI-v2 stayed consistently high across all three scanners.

    Cohort 3 (335 WSIs; eight cases per diagnosis stained at Lab A, each scanned on Aperio GT 450, NanoZoomer S360 and Pannoramic 1000); one classifier per scanner per model, all tested on Aperio-scanned Cohort 1-Test. Accuracies are shown only graphically; no values or tests are given in the text. · 2.3 How robust are foundation models to scanning…; 4.1 WSI datasets; Figure legends

  4. SuggestedWith very few training slides (1 to 8 per diagnosis), the slide encoder TITAN already performed well from a single slide. Patch encoders such as UNI-v2 started lower but improved quickly as slides were added, and all foundation models beat ResNet50.

    Few-shot classifiers trained on 1, 2, 4, 6 and 8 Cohort 3 slides per diagnosis from each scanner, tested on Cohort 1-Test: TITAN exceeded 0.75 accuracy with one case per class; most foundation models exceeded 0.8 accuracy with 8 cases per subtype; best results when the training scanner matched the Aperio test scanner. · 2.4 Do foundation models enable few-shot…; 3 Discussion; Figure legends

  5. SuggestedColour variation followed the scanner more than the staining lab. Stain-vector PCA grouped by scanner, and UNI-v2 tile embeddings formed scanner sub-clusters within each diagnosis. TITAN slide embeddings showed little scanner separation, and neither model grouped by staining lab.

    Vahadane stain vectors from randomly sampled 3000 × 3000 pixel tiles of Cohort 2 slides (stained by 30 labs; Fig. 4 illustrates a synovial sarcoma case): Aperio formed a tight cluster, 3D-Histech was widely dispersed and Hamamatsu was intermediate. t-SNE used three random tumour tiles per slide for UNI-v2 and one embedding per slide for TITAN. Both analyses are qualitative only. · 2.2 Impact of staining and scanning variation; Appendix A Extended Data: Embedding space…; Figure legends

Methods, data and reproducibility

WSIs were tiled with TRIDENT at 20x. Patch embeddings (Path Foundation, UNI-v2, CONCH-v1.5, Virchow, ResNet50) fed an attention-based multi-instance learning classifier, and TITAN/PRISM slide embeddings fed logistic regression. Models diagnosed 14 classes under controlled stain, scanner, mixed-training and few-shot experiments.

  • Cohort 1: H&E WSIs stained at Lab A and scanned on Leica Aperio GT 450 at 40× equivalent; split 60%/20%/20% for train/validation/test · human (patient tumour cases) · n = 1,394 WSIs (approximately 100 cases per diagnosis)
  • Cohort 2: serial 4 µm sections of one representative case per diagnosis, H&E-stained by 30 NHS labs, each slide scanned on three scanners · human (patient tumour cases) · n = 1,243 WSIs
  • Cohort 3: eight cases per diagnosis stained at Lab A, each scanned on Aperio GT 450, Hamamatsu NanoZoomer S360 and 3D-Histech Pannoramic 1000 · human (patient tumour cases) · n = 335 WSIs
  • Cohort 4: real-world archive slides stained at Lab A or 73 other hospitals and scanned on one platform; split into a Combined-cohort training subset and a test subset · human (patient tumour cases) · n = 934 WSIs (121 for Combined training, 813 remaining)
Methods & tools
TRIDENT (tissue detection and tiling), Google Path Foundation, UNI-v2, CONCH-v1.5, TITAN (slide encoder), Virchow, PRISM (slide encoder), ResNet50 CNN (supervised baseline), Attention-based multi-instance learning classifier, scikit-learn logistic regression, Vahadane stain separation with PCA, t-SNE (scikit-learn, Plotly)
Reproducibility gaps
The Data availability section gives a placeholder ('LINK') instead of a URL for the 1,243-WSI multi-institutional release. Sharing of Cohorts 1, 3 and 4 is not described, and no software versions or hardware are reported.
Limitations (6)
  • Only three scanner models (Aperio, Hamamatsu, 3D-Histech) were tested. Devices from other vendors, such as Philips and Roche, were not included, so generalisation to other scanners is not established. · stated by the authors
  • Evaluation covered a single task, soft-tissue tumour subtyping. Other tumour types and tasks, such as mitosis detection or genomic marker prediction, were not tested. · stated by the authors
  • The stain-vector PCA was too widely spread to quantify how much staining and scanning each contribute to colour variation. · stated by the authors
  • Non-Aperio scanned slides were in limited supply, so the scanner experiment trained one classifier per scanner on eight Lab A-stained cases per diagnosis but tested only on Aperio-scanned slides. · stated by the authors
  • Patch encoders used an attention-based MIL head while TITAN and PRISM used unregularised logistic regression. Comparisons between slide and patch encoders, including TITAN's few-shot advantage, therefore mix encoder quality with classifier design, and no repeated few-shot draws are reported. · noted in this summary
  • Cohort 4 counts are internally inconsistent (934 WSIs vs 105 Lab A-stained plus 838 external), and the 813-slide subset is described as training data although the design implies it is Cohort 4-Test. Scanner and few-shot results lack per-model accuracy values or statistical tests in the text. · noted in this summary

Read this if you deploy or fine-tune pathology foundation models across hospitals and need evidence on how H&E stain variation, scanner variation and few-shot training affect slide-level tumour subtyping.

Citations
6
FWCI
20.37
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMeng '26Maleki '25Tada '25Otálora '22Hari '21Laleh '21
64% semantic matchPathology
Yang Meng, Binghao Chai, Tapabrata Chakraborti·bioRxiv·Oct 19, 2026