Chen Hui, Gutt Moritz, Belker Othmar Alexander, Düx Daniel Markus, Wacker Frank K, Hensen Bennet, Gutberlet Marcel
Department of Diagnostic and Interventional Radiology, Hannover Medical School, Hannover, Germany.
Research Campus STIMULATE, Otto-von-Guericke University Magdeburg, Magdeburg, Germany.
Sci Rep. 2025 Aug 25;15(1):31206. doi: 10.1038/s41598-025-11290-6.
In clinical MRI-guided interventions, the lack of high-quality peripheral equipment and specialized interventional MRI systems often necessitates delegating real-time control of MRI scanners to an assistant. We proposed a voice-based interaction system powered by large language models that enabled hands-free natural language control of MRI scanners. The system leveraged multi-agent collaboration driven by large language models to execute scanner functionalities, including sequence execution, parameter adjustment, and scanner table positioning. In 90 hands-free tests for 18 predefined tasks performed within a real MRI scanning room, the system achieved an overall task completion rate of 93.3% (95% CI 86.2-96.9%). On a consumer laptop without GPU support, the response time for control commands was approximately 5-10.5 s. Our study demonstrates the feasibility of using large language models for voice-based interaction with MRI scanners during interventions, eliminating the need for additional assistants and allowing human-like communication.
在临床MRI引导介入中，由于缺乏高质量的外围设备和专门的介入MRI系统，往往需要将MRI扫描仪的实时控制委托给助手。我们提出了一种由大语言模型驱动的基于语音的交互系统，该系统能够实现对MRI扫描仪的免提自然语言控制。该系统利用大语言模型驱动的多智能体协作来执行扫描仪功能，包括序列执行、参数调整和扫描台定位。在真实MRI扫描室内对18项预定义任务进行的90次免提测试中，该系统的总体任务完成率达到93.3%（95%置信区间86.2-96.9%）。在没有GPU支持的消费级笔记本电脑上，控制命令的响应时间约为5-10.5秒。我们的研究证明了在介入过程中使用大语言模型进行与MRI扫描仪的语音交互的可行性，无需额外的助手，并允许进行类人通信。
