A Clinical Benchmark of Foundation Models: Towards Reliable Morphological Subtyping and Cancer Detection on Real-World Barrett’s Esophagus Data

The applicability of emergent histopathology foundation models (Histo-FMs) to real-world diagnostic problems remains unproven. Given the complexity of clinical tasks and the challenges inherent in real-world data, we utilized Histo-FMs to investigate their utility for diagnosing Barrett’s esophagus (BE) and detecting esophageal adenocarcinoma (EAC), a rare malignancy associated with poor patient outcomes. We benchmarked Histo-FMs for these tasks on a real-world cohort representative of routine diagnostics from normal tissue to EAC (N2EAC). The dataset comprised 3,528 hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) from 790 patients (PAXgene-fixed, paraffin-embedded), processed at magnifications ranging from 5× to 40×. A strong multi-rater agreement was achieved between single-scale models for both morphological subtyping and EAC detection. A multi-magnification, multi-backbone aggregation of the five most expert-consistent single-scale models further improved performance (AUROC of 0.907, F1-score of 0.696, accuracy of 0.795, and κ of 0.651 for morphological subtyping; AUROC of 0.909, F1 score of 0.836, accuracy of 0.959, and κ of 0.673 for EAC detection; p<0.05 for most comparisons), indicating robust concordance with expert evaluation. Performance generalized without fixation-specific fine-tuning, underscoring cross-fixation transferability of Histo-FMs. These findings provide the first clinical validation that Histo-FMs can support reliable BE morphological subtyping and EAC detection on real-world data.