Cross-country generalizability of foundation models for cervical cancer screenings on H&E whole slide images

Accurate grading of cervical biopsies on Hematoxylin and Eosin (H&E) stained whole slide images (WSIs) is essential for distinguishing high grade lesions from low grade changes, yet this process is subject to considerable inter-observer variability. In this study, we evaluate a foundation model-based multiple instance learning (MIL) pipeline for binary high-grade squamous intraepithelial lesion (HSIL) detection on H&E stained WSIs. We benchmark our Athena foundation model against four state-of-the-art pathology foundation models: H-optimus-0, Hibou-L, Midnight-12k and Virchow, across datasets from five different countries: Portugal, Cambodia, Germany, Poland and Scotland. Athena achieved the highest mean area under the curve (AUC) (0.931) with the lowest cross-country variability (STD = 0.022). Furthermore, we compared the model's diagnostic performance to that of trained pathologists on a dataset with p16-confirmed ground truth. Our model improved sensitivity from 84% to 95% while maintaining comparable specificity (85% vs. 84%). Failure analysis revealed that the model's errors were concentrated at the diagnostic boundary between low-grade and high-grade lesions, whereas pathologists' errors spanned a broader range of misclassifications. These findings show the potential of foundation models for cervical cancer screenings worldwide.