Foundation models in pathology: bridging AI innovation and clinical practice.

Abstract

Foundation models are revolutionising pathology by leveraging large-scale, pretrained artificial intelligence (AI) systems to enhance diagnostics, automate workflows and expand applications. These models address computational challenges in gigapixel whole-slide images with architectures like GigaPath, enabling state-of-the-art performance in cancer subtyping and biomarker identification by capturing cellular variations and microenvironmental changes. Visual-language models such as CONCH integrate histopathological images with biomedical text, facilitating text-to-image retrieval and classification with minimal fine-tuning, mirroring how pathologists synthesise multimodal information. Open-source foundation models will drive accessibility and innovation, allowing researchers to refine AI systems collaboratively while reducing dependency on proprietary solutions. Combined with decentralised learning approaches like federated and swarm learning, these models enable secure, large-scale training without centralised data sharing, preserving patient confidentiality while improving generalisability across populations. Despite these advancements, challenges remain in ensuring scalability, mitigating bias and aligning AI insights with clinical decision-making. Explainable AI techniques, such as saliency maps and feature attribution, are critical for fostering trust and interpretability. As multimodal integration-combining pathology, radiology and genomics-advances personalised medicine, foundation models stand as a transformative force in computational pathology, bridging the gap between AI innovation and real-world clinical implementation.