Image series analysis represent a common way of analyzing patient condtion by practitioner. Vast amount of parameters and image acquision techniques may lead to various ways of reprsenting volumetric data in a form of image series. Understanding the concepts of image acquisition is important in organization of training sessions for the novice practitioners. For the partcular body part, public providers such as The Cancer Imaging Archive (TCIA) represent a source of vast amount of publicly available collections that cover patient conditions. Although each collection has different focus of research around patient conditions, their studies usually represent image series from arranged exams for patients. Such series are usually accompanied by metadata with unstructured narratives that convey concepts utilized in image acquisition. However, the formatting of such narratives may drastically differ from one collection to the other.

At present, the most recent advances in generative AI results in Large Language models (LLMs) that become a framework of the most AI-powered applications. LLMs such models showcase a promising solution across various fields of natural language processing, including information retrieval (IR).

In this studies we attempt to shed the light on LLMs capabilities in structuring manually written narratives. In the domain of Liver and HCC, using publicly available TCIA collections we construct dataset of precisely structured medical narratives. We conduct an extensive experiments of assessing LLMs on reasoning capabilities in retrieving concepts necessary structuring textual narratives.

Through the results discussion we conclude benefits and limitations of exploiting LLMs in downstream applications: automatic medical practitioner training organizations, medical report assessment, analysis resources content in depth.