A CT Operation Agent "Agentographer" built on Multimodal Large Language Model versus Radiographer in Chest CT Lung Cancer Screening

X-ray Computed Tomography (CT) is widely recognized and globally utilized due to its rapid imaging capability, high spatial resolution, and excellent density resolution. However, the large volume of CT examinations, coupled with complex and repetitive operation procedures, has increased the workload of radiographers and the risk of errors. This is particularly evident in populous countries such as China and India, especially during the prevalence of respiratory infectious diseases or in high-throughout scenarios such as large-scale lung cancer screening. There is an urgent need to enhance the intelligence level of CT examinations to reduce the workload while improving efficiency and quality. To address this challenge, we propose a CT operation agent named Agentographer , which can autonomously perform CT examinations instead of human technicians. We curated a comprehensive multimodal dataset consisting of radiographer-patient interaction videos with associated Q&A transcripts from four distinct brands and models of CT scanners. After that, the dataset was used to train LLaMA-CT, a model specifically fine-tuned for dynamic scene perception and decision-making during CT examinations. Then, powered by the LLaMA-CT, Agentographer integrates AI-driven patient interaction, 3D-camera-based isocenter positioning, segmentation-based CT scan location, automated device control, and AI-driven diagnostic and report models into a unified system. In a subsequent clinical trial of 500 subjects undergoing low-dose chest CT for lung cancer screening, Agentographer independently completed 476 cases (95.2%), achieving a significant reduction in both effective radiation dose (7%, P <0.001) and CT examination duration (9.3%, P <0.001), facilitating radiologists to issue diagnostic reports 12 hours earlier compared to a control cohort of 510 subjects examined by Radiographers.