Cross-Modal Foundation Models for Precision Oncology: Unifying Imaging, Histopathology, Multi-Omics, and Clinical Data

Abstract

Precision oncology has undergone a profound transformation with the emergence of artificial intelligence (AI) capable of integrating heterogeneous biomedical information into unified computational frameworks for individualized cancer care. Traditional machine learning approaches typically analyze single-modality datasets, limiting their ability to capture the complex biological interactions underlying tumor evolution, therapeutic response, and clinical outcomes. Recent advances in cross-modal foundation AI models have fundamentally redefined computational oncology by enabling generalized representation learning across radiological imaging, digital histopathology, genomics, transcriptomics, proteomics, metabolomics, epigenomics, spatial biology, laboratory biomarkers, electronic health records, wearable technologies, and longitudinal clinical outcomes. Built upon transformer architectures, self-supervised learning, graph neural networks, multimodal fusion, and generative AI, these models learn transferable biomedical representations that support diagnosis, biomarker discovery, prognostic prediction, therapeutic optimization, immunotherapy selection, adaptive disease monitoring, digital twins, and intelligent clinical decision support. Emerging technologies including large language models, retrieval-augmented generation, federated learning, reinforcement learning, explainable artificial intelligence, and agentic AI further expand the capabilities of cross-modal computational oncology while accelerating clinical translation. Despite remarkable progress, substantial challenges remain regarding multimodal data harmonization, computational scalability, interpretability, interoperability, regulatory validation, ethical governance, and equitable implementation. This review provides a comprehensive overview of cross-modal foundation models for precision oncology, highlighting computational principles, multimodal integration strategies, clinical applications, implementation challenges, and future perspectives for unifying imaging, histopathology, multi-omics, and clinical data through artificial intelligence.[1]