Coarse-to-Fine Personalized LLM Impressions for Streamlined Radiology Reports

The manual creation of the ”Impression” sectionin radiology reports is a primary driver of radiologistburnout. To address this, we propose acoarse-to-fine framework that uses open-sourceLarge Language Models (LLMs) to automaticallygenerate and personalize impressions from clinicalfindings. The system first generates a draftand then refines it using machine learning andReinforcement Learning from Human Feedback(RLHF) to match individual radiologists’ stylesand ensure factual accuracy. We will fine-tuneLLaMA and Mistral models on a large dataset ofreports from the University of Chicago Medicine.Our approach aims to significantly reduce administrativeworkload and enhance reporting workflowswhile maintaining high standards of clinicalprecision.