Driven by the rapid advancements in artificial intelligence, computational pathology is emerging as a critical engine in the era of precision oncology. Traditional computational pathology primarily relies on task-specific models, which require the development of independent models for each distinct task. This paradigm is heavily dependent on large-scale annotated datasets, resulting in high costs, prolonged development cycles, and poor cross-task adaptability. Consequently, it faces significant challenges in meeting real-world clinical demands for flexibility, accuracy, and broad applicability.
Moreover, task-specific models fall short in addressing the challenges posed by rare diseases, open-set recognition problems, and scenarios with limited sample availability. Against this backdrop, the rise of foundation models (FMs) marks a pivotal shift toward "general intelligence" in computational pathology. Pathology FMs, trained via self-supervised learning on large-scale, multimodal, and multi-center pathology datasets, demonstrate robust representation learning capabilities and generalization performance.
To provide an overview of the recent research advancements in pathological FMs and their applications in oncology, a review study was conducted. It was led by Professor S.Kevin Zhou and Dr. Rui Yan, both from the University of Science and Technology of China, and Dr. Fei Ren from the Institute of Computing Technology, China, along with other collaborators from different institutes. The key challenges and opportunities presented by pathological FMs in precision oncology are also explored. This paper was made available online on September 25, 2025, in the Chinese Medical Journal
These models enable multi-task transfer with minimal or even zero annotated data, significantly enhancing clinical utility and generalizability. Current research on pathology foundation models can be broadly categorized into three types:
Pathology Image Foundation Models: These models focus on extracting critical visual features from whole slide images (WSIs), and are applicable to tasks such as cancer type classification, tumor grading, and biomarker prediction. Representative models, including GigaPath, UNI, and Virchow, have achieved performance surpassing conventional approaches across various cancer types.
Pathology Image-Text Foundation Models: By integrating natural language processing capabilities, these models are jointly pre-trained on pathology images and textual reports. They support tasks such as cross-modal retrieval, diagnostic report generation, and educational assistance in pathology. Representative models include PLIP, CONCH, and PathChat. Notably, these models enable zero-shot learning, making them applicable to previously unseen case types and supporting digital pathology education and remote diagnosis. Unlike traditional pathology models, PLIP integrates natural language annotations, which enables it to comprehend image semantics, and perform various downstream tasks
Pathology Image-Gene Foundation Models: These models align and model pathology images alongside multi-omics data, such as transcriptomics and proteomics. Notable examples like mSTAR, GiMP, and TANGLE have not only improved the precision of tumor subtype classification and treatment response prediction but also provided new avenues for understanding cancer heterogeneity and underlying molecular mechanisms.
Collectively, these foundation models constitute a new cornerstone in computational pathology infrastructure and expand the boundaries of digital pathology applications. In clinical applications, foundation models significantly increase the efficiency and accuracy of pathological analysis and reduce the reliance on specialized and costly techniques such as IHC and genomic testing. This not only accelerates diagnostic processes but also lowers health care costs and improves patient treatment experiences. Despite their promising performance across various domains, pathology foundation models still face several challenges in clinical deployment. First, most existing models lack extensive validation on multi-center, real-world datasets, limiting their reliability, stability, and fairness in complex healthcare settings. Second, the "black-box" nature of these models remains a major barrier to clinical trust. Enhancing interpretability and transparency regarding biological mechanisms has thus become a key research focus. Additionally, in multimodal integration tasks, issues such as feature redundancy, informational conflict, and data heterogeneity between modalities remain unresolved.
Future research must advance in areas including long-sequence modeling, high-dimensional feature fusion, ethical guideline development, and intelligent cross-modal collaboration. Foundation models are not only reshaping the research paradigm of computational pathology but also laying the technical groundwork for intelligent, automated, and personalized pathology decision-support systems. Their continued evolution promises to drive transformative change in precision oncology and life science research.
