Evaluating GPT4 on Impressions Generation in Radiology Reports

Generation of Impressions from radiology report Findings is a critical aspect of medical image analysis, assisting clinicians in making informed decisions (1).Traditionally, this process requires manual input from the interpreting radiologist, which can be time consuming and occasionally can be inconsistent with the Findings section.Fine-tuned pretrained models have shown promise in automating or proofreading this task (2); however, they often necessitate substantial training data sets, which may not always be accessible in specialized domains, such as radiology.The recent success of large language models, such