Multimodal Cancer Foundation Models: Bridging Radiology, Pathology, Genomics, and Electronic Health Records

Abstract

Cancer remains one of the leading causes of morbidity and mortality worldwide despite remarkable advances in molecular diagnostics, targeted therapeutics, immunotherapy, and precision medicine. The rapid expansion of biomedical data generated through radiological imaging, digital pathology, genomic sequencing, transcriptomics, proteomics, metabolomics, laboratory investigations, wearable technologies, and electronic health records has created unprecedented opportunities for artificial intelligence (AI)-driven precision oncology while simultaneously introducing major computational challenges. Multimodal cancer foundation models have emerged as a transformative computational paradigm capable of learning generalized biomedical representations across heterogeneous data modalities using large-scale self-supervised learning. Unlike conventional machine learning systems designed for individual diagnostic tasks, these models integrate radiology, pathology, genomics, laboratory biomarkers, longitudinal clinical information, and electronic health records into unified patient-centered computational representations that support diagnosis, prognostic prediction, molecular characterization, therapeutic optimization, clinical decision support, and personalized medicine. Advances in transformer architectures, multimodal learning, graph neural networks, large language models, retrieval-augmented generation, foundation model pretraining, and generative AI have substantially accelerated development of multimodal oncology ecosystems capable of comprehensive biomedical reasoning across the entire clinical workflow. Despite remarkable progress, important challenges remain regarding multimodal data harmonization, computational scalability, explainability, interoperability, cybersecurity, regulatory validation, and equitable implementation. This review provides acomprehensive overview of multimodal cancer foundation models, highlighting computational principles, clinical applications, implementation challenges, and future perspectives for bridging radiology, pathology, genomics, and electronic health records in precision oncology.[1]