Hao Cheng, Chi Chun Wong
The integration of Artificial Intelligence (AI) into cancer pathology offers an imperative solution to global pathologist shortages and increasingly complex diagnostic demands. This review summarized the rapid evolution of AI in the field, highlighting the paradigm shift from task-specific (TS) algorithms towards powerful, versatile foundation models (FMs), such as UNI, CONCH, GigaPath, mSTAR, and Atlas. These models, trained on massive and diverse datasets using self-supervised and multimodal learning, demonstrate remarkable capabilities in cancer classification, subtyping, outcome prediction, and biomarker discovery. The emergence of AI "copilots", such as PathChat, SmartPath, further promises to streamline workflows through conversational interfaces and autonomous task planning. However, significant challenges impede clinical translation, including a validation crisis underscored by poor generalizability in zero-shot testing, critical concerns regarding model explainability ("black-box" nature), risks of hallucinations in generative tools, and ensuring generalizability and fairness across diverse populations. Robust external validation, standardized benchmarking, development of explainable AI approaches, and novel regulatory frameworks are essential to responsibly harness the transformative potential of foundation models and realize their promise in improving diagnostic accuracy, efficiency, and patient outcomes in cancer pathology.