Foundation Models in Cancer Pathology: Techniques, Applications, and Future Directions

Computational pathology enables scalable analysis of pathology images for cancer diagnosis and research, but conventional deep learning models remain constrained by annotation dependence and task-specific development. Recent advances in self-supervised learning, transformer-based architectures, and large-scale pretraining have given rise to computational pathology foundation models (CPathFMs), which are designed to learn transferable and reusable representations from diverse pathology data. This review first traces the development of CPathFMs by examining the data resources, backbone architectures, representation levels, input modalities, and pretraining strategies that shape their design. We then examine their applications across major cancer pathology tasks, including tumor detection, grading, and subtyping, molecular biomarker and gene expression prediction, prognostic assessment, tissue phenotyping, as well as multimodal retrieval and report generation. We further discuss key challenges that limit translation and deployment, including underrepresentation of normal tissues and benign lesions, limited generalization across domains and institutions, data overlap and benchmark contamination, task-centric evaluation and insufficiently standardized metrics, high training and infrastructure costs, and ethical concerns in model governance. Overall, CPathFMs mark an important shift from task-specific pathology artificial intelligence toward reusable representation learning. Their clinical value remains to be established through improved model performance and generalizability, more comprehensive and standardized evaluation frameworks, and prospective evidence of utility in real-world pathology workflows.