Can LLMs Turn French PET/CT Narrative Reports into Structured Knowledge?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Natural Language Processing

Background:

• Cognitive impairment diagnosis relies on interpreting complex PET/CT reports.
• Information extraction from clinical text is challenging but crucial for research and patient care.

Purpose of the Study:

• To evaluate the efficacy of large language models (LLMs) for extracting descriptive and interpretative patterns from French PET/CT reports in cognitive impairment.
• To compare the performance of a large multilingual LLM (GPT-OSS) and a smaller specialized model (NuExtract 2.0).

Main Methods:

• Utilized a corpus of 620 annotated French PET/CT reports from Geneva University Hospitals.
• Applied two open-weight LLMs (GPT-OSS 120B and NuExtract 2.0 8B) in zero- and few-shot settings.
• Employed clustering-based shot selection for few-shot learning.

Main Results:

• GPT-OSS (120B) demonstrated superior accuracy in information extraction compared to NuExtract 2.0 (8B).
• The larger GPT-OSS model required significantly more computational time (6x).
• Feasibility of using multilingual LLMs for French clinical narrative analysis was supported.

Conclusions:

• Multilingual LLMs, particularly larger models (120B) with few-shot examples, are effective for extracting information from French PET/CT reports.
• Further studies involving fine-tuning are warranted to confirm these findings.
• The approach shows potential for extending to automated diagnosis prediction in cognitive impairment.