Designing MR Exams Using an Autonomous Multi-Agent Large Language Model System

Motivation: There is a global shortage of radiology staff which can lead to workforce burnout, increased backlogs and healthcare costs. AI-based workflow automation can help mitigate some of these effects. Goal(s): To develop an autonomous large language model system that can generate MR pulse sequences customized to the patient's electronic health record. Approach: We developed a multi-agent system that takes natural language prompts and produces scanner-executable Pulseq sequences via EHR database queries and customized MR protocol generation. The system was validated using the MIMIC-IV database. Results: The multi-agent system successfully generated patient-specific pulse sequences that achieved desired tissue contrast in brain imaging. Impact: MR exam delivery is challenged by a worldwide shortage of radiology staff. We demonstrate that a multi-agent LLM system shows promise in automating MR exams by accessing a patient's health record and designing the protocol and sequences to be acquired.