AI-assisted extraction of opportunistic findings from oncologic CT reports using large language models.
Authors
Affiliations (10)
Affiliations (10)
- University of Udine, Udine, Italy. [email protected].
- Department of Medicine (DMED), Institute of Radiology, University of Udine, Udine, 33100, Italy.
- University of Udine, Udine, Italy.
- Department of Radiology, Campostaggia Hospital, Azienda USL Toscana Sud-Est, Poggibonsi, 53036, Italy.
- Department of Medical Oncology, CRO di Aviano, National Cancer Institute, IRCCS, Aviano, 33081, Italy.
- Department of Medicine, University of Udine, Udine, 33100, Italy.
- Institute of Radiology, University Hospital S. Maria della Misericordia, Azienda Sanitaria Universitaria Friuli Centrale (ASUFC), Udine, 33100, Italy.
- University of Udine, Udine, Italy. [email protected].
- Department of Medicine (DMED), Institute of Radiology, University of Udine, Udine, 33100, Italy. [email protected].
- Institute of Radiology, University Hospital S. Maria della Misericordia, Azienda Sanitaria Universitaria Friuli Centrale (ASUFC), Udine, 33100, Italy. [email protected].
Abstract
Opportunistic findings at imaging (iOFs), such as osteoporosis, liver steatosis, or coronary artery calcifications, are clinically relevant abnormalities that may be visible on routine computed tomography (CT) examinations but are often underreported in oncologic imaging. This underreporting reflects high radiologist workload, limited awareness of secondary findings, and the narrative structure of radiology reports. In this study, we combine clinical evaluation with artificial intelligence to investigate the detection and reporting of iOFs in routine oncologic CT examinations. In a multi-phase study involving 156 oncology patients and five radiologists, we assess inter-reader agreement for ten predefined iOFs, observing high concordance across findings and reader experience levels. We then develop a Large Language Model (LLM)-based system to automatically extract iOFs from free-text radiology reports and link them to predefined, guideline-based management recommendations. The system shows high agreement with report-level expert annotations of the information explicitly documented in free-text reports (F1 = 0.995 in internal testing and 0.980 in independent reader validation). Retrospective analysis identified previously unknown opportunistic findings without documented dedicated follow-up in 29 of 156 patients (18.6%), supporting the potential clinical value of systematic extraction and structured communication. These results suggest that LLM-based analysis of radiology reports may support the systematic extraction and communication of documented opportunistic findings in clinical workflows.