The emergence of diverse pathological foundation models has empowered computational pathology tasks within the field of oncology, including tumor subtyping, cancer prognosis, biomarker prediction, gene expression prediction, and so on. However, variations in model architecture and data sources hinder consistent downstream performance and complicate centralized training. Specifically, the lack of data sharing makes retraining foundation models with pooled data infeasible. Here, we propose the Meta-Encoder, a unified framework that integrates features from multiple pathological foundation models to generate a comprehensive representation, achieving superior performance in downstream cancer detection tasks compared to single foundation models. While single models suffice for low-complexity univariate tasks such as cancer diagnosis and prognosis, the Meta-Encoder consistently rivals the best-performing single model, alleviating concerns over model selection. For high-dimensional tasks such as multiplex protein and gene expression prediction within tumor tissues, the attention-based strategies of our Meta-Encoder framework demonstrate substantial advantages over single models, offering an optimal balance of performance and efficiency. By harnessing the complementary strengths of multiple foundation models, the Meta-Encoder enhance the molecular characterization of pathology images, thereby advancing precision oncology.
Gao et al. (Wed,) studied this question.