Generating Personalized MRI Reports from Key Phrases with Multiple LLM Agents and RAG Technique

抄録

In the paper, the novel method for generating personalized magnetic resonance imaging (MRI) reports based on key phrases is introduced. In order to demonstrate the method, an application with multiple large language model (LLM) agents was written. For establishing connection between different agents, graph-based software architecture was used. The graph consists of the following nodes. (1) The data retriever node. In this node relevant reports along with the metadata are obtained from the vector database. (2) The findings generation node. This node includes an implementation of two different variations of an MRI report’s findings section generation method. The first is based on merging already existed findings parts of radiology reports that were retrieved by key phrases into the new one. The second is based on reconstructing findings section so that it contains specific paragraphs retrieved by key phrases from the vector store. (3) The impression generation node. This node implements summarization of a findings section into an impression. The performance of a findings part generation was evaluated with BLEU, ROUGE and BERTScore metrics. Experiments have shown promising results in generation of findings parts: for the first approach mean ROUGE-L score was about 0.5, BERTScore was around 0.8, for the second approach mean ROUGE-L score was approximately 0.6, BERTScore was roughly 0.9. The first approach was around three times slower than the second. During the experiments, the dataset of 908 reports gathered from seven radiologists was used. Qualitative analysis was performed with five-point Likert scale questionary and statistically analyzed by means of Mann–Whitney U test.