Using Large Language Models and Retrieval-Augmented Generation in MRI Protocol Selection: Balancing Accuracy and Privacy

Motivation: When medical staff lack clinical experience, they might pick the wrong protocols and parameters, leading to wasted time and resources. Goal(s): We want to show that combining RAG technology with LLMs can help determine MRI protocols at a level matching seasoned professionals —all while keeping patient privacy intact. Approach: Using clinical doctors' MRI exam requests as our benchmark, we compared accuracy across different experience levels and professions to see how cloud-based and local LLMs differ in performance. Results: After adding RAG technology, the cloud-based LLM matched the expertise of experienced radiologists, while the local LLM reached accuracy similar to that of senior radiologic technologists. Impact: We've proven that RAG-based LLMs are feasible for early MRI decision-making, offering a new tool for learning and error prevention. Cloud-based LLMs and local LLMs each have their strengths in accuracy and privacy, but neither is perfect just yet.