Can LLMs Turn French PET/CT Narrative Reports into Structured Knowledge?
Authors
Affiliations (4)
Affiliations (4)
- Division of Medical Information Sciences, Geneva University Hospitals, Switzerland.
- Department of Radiology and Medical Informatics, University of Geneva, Switzerland.
- Div. of Nuclear Medicine and Molecular Imaging, Geneva Univ. Hosp., Switzerland.
- NIMTLab, Fac. of Medicine, U of Geneva, Center of biomedical imaging, Switzerland.
Abstract
This study evaluates large language models (LLMs) for information extraction from French PET/CT reports related to cognitive impairment, focusing on descriptive patterns of cerebral metabolism, perfusion and uptake for three radiotracers, as well as interpretative patterns (Braak stage, positivity and diagnosis). A corpus of 620 annotated reports from the Geneva University Hospitals was used to test two recent open-weight models: GPT-OSS (120B), a multilingual generalist model, and NuExtract 2.0 (8B), smaller but specialized in structured data extraction. Both were applied in zero- and few-shot settings using a clustering-based shot selection. GPT-OSS achieved superior accuracy but required 6 times more computation time. Results support the feasibility of applying multilingual LLMs to French clinical narratives, preferably using a larger model (120B) and few-shot examples. These findings warrant further confirmation studies with fine-tuning and encourage extending the approach to diagnosis prediction.