From cells to patients: Multiscale computational pathology in the era of foundation models and vision-language systems

Computational pathology is fundamentally defined by its inherent hierarchical structure, spanning from nuclear morphology and cellular interactions to tissue microenvironments, ultimately integrating into whole-slide images for patient-level prognostic profiling. While traditional deep learning approaches have achieved remarkable success in specific tasks, the recent emergence of large-scale foundation models and vision-language models has precipitated a paradigm shift in the field. These data-driven systems, characterized by their robust representation learning and semantic reasoning capabilities, are redefining how we analyze pathological data across diverse spatial scales. In this review, we provide a comprehensive synthesis of this transformation through a multiscale lens. We systematically survey the application of foundation models and vision-language models in deciphering biological complexity, ranging from cell-level segmentation and tissue phenotyping to slide-level prediction and multimodal integration. Furthermore, we critically analyze the limitations of current approaches, such as interpretability, computational efficiency, and data bias, then outline promising future directions for developing holistic, context-aware systems that bridge the gap between pixel-level features and patient-centric clinical decision-making.