Foundation model performance across multi-institutional whole-slide image cohorts: a validation study

Berg L, Okonkwo A, Yamada H, Fernández M

Artificial Intelligence in Diagnostics Journal · Published 15 Jul 2026 · DOI: 10.00000/aidj.2026.0143

Abstract

Foundation models trained on large unlabelled whole-slide image corpora have shown strong transfer performance, but external validity across scanners, staining protocols and populations remains incompletely characterised. We evaluated three publicly available pathology foundation models across seven institutional cohorts comprising 41,300 whole-slide images, assessing diagnostic classification, site-of-origin prediction and biomarker inference. Performance degraded measurably under scanner shift and was partially recovered by stain normalisation and site-stratified calibration. We provide a reporting checklist for external validation of pathology foundation models.

Keywords

Full text

Introduction

The full text of this article is available in the production system. This demonstration build presents the abstract, metadata, figures and reference structure to illustrate the article page layout.

Methods

Study design, data sources, inclusion criteria, statistical analysis and ethical approval are reported in full in the published version.

Results

Figures, tables and supplementary files accompany the published article.

Discussion

Interpretation, limitations and implications for practice are presented in the published version.

References

1. Reference list is provided with the published article.

Cite this article