Cross-Organ Standalone Performance of a Gastric-Trained Artificial Intelligence Algorithm in Esophageal Biopsies

Presented at ECDP 2026

Authors

Giuseppe Mallel¹, Anastasia Mercurio², Shaya Goodman¹, Inbal Gross¹, Manuela Vecsler¹, Emanuela Pilozzi³

Affiliations

1. Ibex Medical Analytics, Tel Aviv, Israel
2. Department of Experimental Medicine, Sapienza University of Rome, Rome, Italy
3. Department of Clinical and Molecular Medicine, Sapienza University of Rome, Rome, Italy

Introduction

Barrett’s esophagus (BE) is the main precursor to esophageal and gastroesophageal junction (GEJ) adenocarcinoma. Diagnosing BE-related dysplasia can be challenging with high interobserver variability. We assessed an artificial intelligence (AI) tool, trained solely on gastric whole slide images (WSIs) to detect relevant morphologies in esophageal biopsies.

Materials and Methods

A retrospective dataset of 425 H&E-stained WSIs from 240 esophageal and GEJ biopsies was collected from a single institution. The cohort included 82 BE (10 dysplasia/adenocarcinoma), 75 eosinophilic esophagitis, 33 GERD-related esophagitis, and 50 normal cases. All slides were scanned at 40x (NanoZoomer S360). The standalone performance of the algorithm (Ibex Gastric) was assessed against an established ground truth.

Results

The AI algorithm showed high performance across the three evaluated esophageal morphologies. For BE (intestinal metaplasia), the AI achieved an AUC of 0.9924 (0.987–0.998) with a sensitivity of 96.4% (91.1–98.6%) and a specificity of 97.4% (94.6–98.7%). In detecting Low-Grade Dysplasia, the AI demonstrated an AUC of 0.984 (0.964–1.000) (sensitivity 100% (72.2–100%), specificity 96.0% (90.2–98.4%)). Finally, for High-Grade Dysplasia/Adenocarcinoma, the AUC was 0.927 (0.855–1.000) (sensitivity 100% (51.0–100%), specificity 86.4% (78.7–91.6%)).

Conclusions

The AI algorithm, trained only on gastric tissue, effectively detected BE and BE-related dysplasia in esophageal biopsies. This cross-organ generalization, despite possible confounders such as squamous epithelium and esophagitis, suggests that the AI learns generalizable morphological features. Following further validation, this AI could be a valuable decision-support tool for pathologists, reducing variability and increasing confidence in esophageal diagnoses.