S943 Automated Detection of Barrett's Esophagus Dysplasia Using Deep Learning With Pathology Foundation Models on a Large Multicenter Cohort of Whole Slide Images

Introduction: The histologic diagnosis of dysplastic Barrett's esophagus (BE) suffers from substantial interobserver variability due to subjective criteria and inflammatory artifacts, leading to dysplastic overcalls particularly in the community setting. Precise dysplasia grading is critical for determining appropriate management. We aimed to develop and validate a deep learning model using a large multi-center cohort to improve the accuracy of BE dysplasia grade diagnosis. Methods: We utilized 969 whole slide images (WSIs) from 4 academic centers (Mayo Clinic, Baylor Scott and White Medical Center, Oregon Health and Science University, and University of Alabama) collected from 1992 to 2022. All patients had endoscopic BE (≥1 cm columnar mucosa) with confirmed intestinal metaplasia. Ground truth was established by consensus review by 2 expert gastroenterology pathologists. Data were stratified by dysplasia grade and center and split 70-15-15 for training, validation, and testing. We employed a semi-supervised approach using Clustering-constrained Attention Multiple Instance Learning (CLAM) architecture. WSIs were divided into patches with features extracted using the Midnight foundation model, a pre-trained deep learning model that learned to identify histologic patterns from large datasets of pathology images. Three binary classifiers identified candidate patches followed by the final CLAM model predicting WSI-level dysplasia grade. Model performance was evaluated using area under the curve (AUC) and F1 score, a balanced measure of diagnostic accuracy accounting for both correct identification of dysplasia and avoidance of false positives. Results: The mean age was 64.6 ± 12.8 years and 81.9% were men. Dysplasia grade by consensus diagnosis was 365 nondysplastic BE (NDBE), 319 low-grade dysplasia, and 285 high-grade dysplasia. On the standalone test set, the model achieved AUC of 95%, F1 score of 82%, sensitivity of 82%, and specificity of 91% for 3-class grading. For binary classification of any dysplasia versus NDBE, AUC was 97%, F1 score was 93%, sensitivity of 90%, and specificity of 94%. Conclusion: Our model demonstrates promising performance for BE dysplasia diagnosis grade. The combination of a large multi-center training dataset and foundation model training provides inherent robustness accounting for staining variations across institutions without requiring center-specific adaptations. This approach may serve as a valuable adjunct to pathologist review, improving diagnostic consistency. External validation of this model is planned.