.png){kind=icon}
.png){kind=icon}
Return
An LLM-Based Active Assistant and Smart Manual for CT Imaging Workflows
.png){kind=icon}
PRE
AI
DOI：10.1007/978-3-032-07502-4_6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Performing medical imaging exams ideally requires technologists to have deep clinical knowledge and technical expertise to operate scanners effectively and efficiently. However, there is a shortage of highly qualified workforce in radiology while patient volumes are rising steadily. Therefore, intelligent assistance is needed to streamline workflow adjustments and reduce manual workload. This study presents a Large Language Model (LLM)-based chatbot to assist clinical staff in Computed Tomography (CT) protocol and postprocessing setup by integrating device-specific information and patient-specific clinical data. As an active assistant integrated in a scan workflow prototype, it provides responses with actionable links that allow users to modify protocol settings directly. At the same time, it acts as a smart manual enhanced with patient-specific context, referencing and linking to both official device documentation as well as clinical indication and prior diagnostic reports. This is realized with an advanced Retrieval Augmented Generation (RAG) with pre- and post-retrieval strategies to improve contextual relevance. An LLM-based evaluation was employed to assess performance. We achieved 95.0% alignment with predefined expectations using GPT-4o mini, and 98.3% with GPT-4o. To evaluate the effect of the applied techniques, an ablation study was conducted. Omitting few-shot examples and instruction-based prompting reduced expectation alignment to 71.4% and 60.5%, respectively. When both were removed, it decreased to 55.0%. The findings underscore the effectiveness of prompt engineering in guiding LLMs to produce accurate, clinically relevant outputs in the correct format.
Keywords:
CT workflow assistant
LLM
Prompt engineering techniques
Journal
E
IF:
0
Papers:
14
Citations:
0