Predicting MRI Protocol Using an Adapted Large Language Model

Motivation: Developing a system to predict radiology protocols for incoming MRI orders. Goal(s): Optimizing protocol selections by developing a decision-support system, enhancing efficiency and accuracy in selecting protocol elements Approach: We adapted an open-source LLM using parameter-efficient fine tuning to ingest free-text inputs generated from ordering physicians to predict four key elements of MRI protocols: broad anatomical region, the focus target organ, contrast use, and protocol title. Results: The model achieved high F1-scores (approximately 95%) for Region and Contrast predictions, while Focus and Protocol metrics were lower due to missing labels and broad anatomical terms, challenging accurate predictions. Impact: The proposed system would offer radiologists a privacy-preserving decision-support tool, potentially reducing protocol mismatches, enhancing diagnostic accuracy, and optimizing workflow. Streamlining MRI protocoling aims to enhance diagnostic quality, safeguard patient health, expedite treatment, and lower healthcare costs.