The Current Status of AI-accelerated MRI Techniques in Clinical Use

作者：Sven Haller, Dennis M. Hedderich, Christian Federau, Christian Weisstanner, Myriam Edjlali, Sofie Van Cauter, Greg Zaharchuk · 发表于：Radiology · 年份：2025 · DOI：10.1148/radiol.243819 · 被引用次数：10 · 研究领域：Advanced MRI Techniques and Applications、Atomic and Subatomic Physics Research、Radiology practices and education

Artificial intelligence (AI) tools to accelerate MRI are rapidly entering clinical routine. Several techniques for MRI acceleration already exist, including compressed sensing and parallel imaging. The introduction of AI acceleration tools for MRI is therefore not fundamentally novel. However, the possibility of combining these AI tools with existing MRI acceleration techniques adds potential opportunities and complexity. This article focuses on commercially available AI tools for clinical MRI acceleration. The basic principle of AI-accelerated MRI is to shorten acquisition time-which results in noisier or lower-spatial-resolution images-then recover image quality with AI. The potential advantages of AI-accelerated MRI include increased patient comfort, shorter waiting lists, reduced motion artifacts, economic efficiencies, and environmental benefits. This article first briefly presents fundamental technical aspects of AI acceleration tools, including noise reduction and super-resolution reconstruction, summarizing available evidence. Potential errors and pitfalls, notably hallucinations (ie, invented or disappearing lesions) are serious concerns, yet they remain poorly investigated. The occurrence of hallucinations, however, is probably rare at the acceleration levels recommended for clinical practice. The downstream implications and potential challenges of AI-accelerated MRI are also discussed, including generating too many images and studies for a limited number of radiolo...