EBV status prediction in gastric carcinoma biopsy images using multiple bag instance learning and foundation model in multicenter data

Epstein-Barr virus associated gastric carcinoma (EBVaGC) represents a distinct molecular subtype of gastric cancer with prognostic and therapeutic significance. Current diagnostic practice relies on Epstein-Barr virus-encoded small RNA (EBER) in situ hybridization, which is rarely performed on initial biopsy specimens. We developed a deep learning-based diagnostic pipeline that predicts EBV status directly from H&E-stained gastric biopsy slides. Our method incorporates a multiple bag instance learning (MBIL) framework, which groups spatially adjacent tumor patches to address the inherent heterogeneity of biopsy samples and improve predictive performance. A pathology foundation model was employed to distinguish tumor from normal regions with minimal pathologist annotation, enabling robust tumor detection and contributing to downstream model accuracy. The model was trained on 744 biopsy cases from three general hospitals using 5-fold cross-validation and evaluated on an independent external cohort of 157 cases. MBIL-based models consistently outperformed single-bag approaches. The DSMIL-MB model achieved the best performance, with an area under the receiver operating characteristic curve (AUC) of 0.907 ± 0.044 in internal validation and 0.870 in external testing. TransMIL also showed improved performance with MBIL, with external AUC increasing from 0.701 to 0.853. This framework may serve as a practical prescreening tool for EBV status prediction in routine gastric biopsy specimens prior to EBER-ISH testing, supporting preoperative decision-making and facilitating broader implementation of EBV screening in gastric cancer diagnostics. © 2026 The Pathological Society of Great Britain and Ireland.