Adapting pathology foundation models for continual cross-center WSI retrieval.

The construction of medical centers is rapidly advancing, generating a vast amount of whole slide images (WSIs). Content-based histopathological image retrieval (CBHIR) unlocks the rich digital morphologic content of WSIs previously confined to glass slides. Foundation models trained on large-scale pathology data have shown remarkable generalization and transfer capabilities, providing a powerful basis for CBHIR. However, deploying pathology foundation models across different centers remains challenging due to cross-center domain shifts and continual data expansion, which can lead to feature drift during long-term model adaptation. To address these issues, we present a continual learning framework that adapts pathology foundation models for continual cross-center WSI retrieval (CCBHIR). Our framework aligns outputs of pre-trained pathology foundation models into a unified latent domain by generating instance-wise prompts that dynamically mitigate domain discrepancies. In addition, an embedding consistency replay mechanism enables stable and efficient feature rehearsal without rebuilding the entire index, thus preserving both forward and backward retrieval compatibility across centers. Evaluations on a large-scale continual retrieval dataset comprising 10,837 WSIs from TCGA projects demonstrate that our framework achieves superior intra-center and cross-center retrieval performance compared with state-of-the-art continual learning methods. This work provides an effective strategy for adapting pathology foundation models to real-world, multi-center deployment scenarios, bridging the gap between foundation model research and practical computational pathology applications. The code is available at https://github.com/OliverZXY/CCBHIR.