Foundation models in clinical oncology: Progresses and perspectives

Cancer remains a leading global burden, and the volume of multimodal oncology data-spanning digital pathology, radiology, genomics, and Electronic Health Records (EHRs)-is expanding exponentially. Foundation models (FMs), trained on broad unlabeled corpora and adapted to diverse tasks, have catalyzed advances in screening and diagnosis, outcome prediction, and inference of actionable molecular alterations, with early traction in treatment selection, drug repurposing, and clinical trial design. Yet key barriers persist: variability in tissue processing, imaging protocols, and data standards drives heterogeneity; the scarcity of well-annotated, multi-institutional cohorts limits development and independent validation; regulatory guidelines and policy making for FM-enabled diagnostics are still maturing; privacy and data sovereignty complicate data sharing; and bias against under-represented populations threatens equitable performance. A practical way forward is to combine high-quality, representative cohorts and interoperable data infrastructure with cross-institutional partnerships that embed AI talent within clinical teams-closing these gaps and enabling generalizable, trustworthy tools to reshape cancer care.