Acad Radiol. 2025 Nov 19. pii: S1076-6332(25)01049-9. [Epub ahead of print]
Foundation Models (FMs) mark a significant evolution in medical AI, enabling multimodal and multitask performance across text and imaging. Radiology, with its structured data formats and early adoption of AI, is uniquely positioned to benefit from FM capabilities. However, despite promising technical advances, questions remain about their clinical readiness, safety, and regulatory oversight. This narrative review explores the development, utility, and implementation challenges of FMs in U.S. radiology. Literature from PubMed, Scopus, arXiv, and IEEE Xplore (January 2022 to May 2025) was synthesized to highlight architectural trends, clinical applications, evaluation methods, and regulatory developments. U.S.-based models like CheXzero, BioMedCLIP, and Med-PaLM demonstrate strong diagnostic and reporting performance but face key limitations-including lack of FDA clearance, limited external validation, and integration barriers with PACS/RIS systems. Safety issues such as hallucination, automation bias, and underperformance in edge cases persist. While human-in-the-loop frameworks, federated learning, and emerging reporting standards show promise, institutional readiness and regulatory clarity remain fragmented. We propose a roadmap that includes continuous monitoring, equity-focused design, and a national FM registry to guide responsible deployment. Radiology's digital maturity makes it a critical testbed for foundational AI integration-offering lessons for broader clinical adoption.
Keywords: Bias and Health Equity; Clinical Validation; Diagnostic Imaging AI; FDA Regulation; Federated Learning; Foundation models; Human-in-the-Loop AI; Multimodal Learning; Radiology Artificial Intelligence; Vision-Language Models