Enhancing LLMs for Impression Generation in Radiology Reports through a Multi-Agent System

In radiology workflow, radiologists traditionally interpret imaging studies and manually draft detailed reports, including an "impression" section that summarizes clinically significant findings and possible diagnosis, which is a vital part of the report for referring physicians and patient care. This process is time-consuming and subject to variability impacted by the radiologist's knowledge and experience [1]. Automated impression generation has the potential to improve report consistency, reduce radiologist workload, and enhance the overall quality of radiology reports [2]. Such a feature is especially needed with the recent growth in the demands for medical imaging, which are straining radiologists, leading to possible burnout and impacting their ability to provide timely and precise reports [3]. Large Language Models (LLMs) have shown exceptional capabilities in understanding and generating text that is coherent and contextually relevant, making them promising tools for auto-generating impressions from findings in radiology reports. A few studies have investigated LLMs' ability for the impression summarization task [2, 4, 5, 6, 7, 8], demonstrating the potential of LLMs to revolutionize radiology workflow by automating the report generation process. Various techniques, such as prompt engineering, model fine-tuning, and retrieval-augmented generation (RAG) [6], have been used to improve their performance and mitigate the limitations of LLMs, including hallucinations and insufficient domain knowledge. However, current approaches often lack the ability to integrate these techniques together within a unified framework and the scalability to incorporate new techniques and/or requirements in impression generation.