Foundation Model Ensemble for Out-of-Distribution Generalization: Predicting Lymph Node Metastasis in Early Gastric Cancer Using Whole-Slide Imaging

Abstract

Recent advances in deep learning have improved the practicality of automated analysis for whole-slide imaging. However, challenges remain in image analysis due to variations in imaging equipment, tissue preparation, staining protocols, and other variables. These variations hinder the generalizability of trained models to external datasets. Recently, foundation models trained on large-scale pathology datasets have been introduced by various research groups, demonstrating the potential to address this issue. Since each foundation model was trained on datasets collected from different sources under varying settings, the learned representations reflect different characteristics to some extent. These differences suggest that leveraging the information of multiple models could improve generalization and robustness compared to using a single model. In this study, we investigate foundation model ensembles for predicting lymph node metastasis in early gastric cancer across three different datasets. By comparing ensemble models with individual ones, we demonstrate that ensembling multiple foundation models improves performance in whole-slide imaging for both in-distribution and out-of-distribution data.