Automated Structured Reporting of Neurodegenerative Imaging Markers Using Quantitative MRI and Large Language Models.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea (J.S.K.).
- VUNO Inc, Seoul, Republic of Korea (W.J., D.-H.K., S.H.L., M.K., N.L., S.K.); UC Berkeley-UCSF Graduate Program in Bioengineering, University of California, San Francisco, San Francisco, California (W.J.).
- Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yonsei University College of Medicine, Seoul, Republic of Korea (P.S.S.). Electronic address: [email protected].
- Department of Medical Science, Asan Medical Institute of Convergence Science and Technology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Republic of Korea (H.H., W.H.S.).
- Department of Medical Science, Asan Medical Institute of Convergence Science and Technology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Republic of Korea (H.H., W.H.S.); Department of Radiology and Research Institute of Radiology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Republic of Korea (W.H.S., S.J.K., C.H.S.).
- VUNO Inc, Seoul, Republic of Korea (W.J., D.-H.K., S.H.L., M.K., N.L., S.K.).
- Department of Radiology and Research Institute of Radiology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Republic of Korea (W.H.S., S.J.K., C.H.S.).
Abstract
To evaluate the feasibility of a quantitative magnetic resonance imaging (MRI)-to-structured reporting pipeline that integrates quantitative MRI software with large language models (LLMs) for structured drafting of neurodegenerative imaging markers and generating clinician-facing report conclusions. A total of 192 patients with 3D T1 and 2D or 3D FLAIR MRI images from the Alzheimer's Disease Neuroimaging Initiative database were randomly selected. A deep-learning model generated structured report findings based on nine indices related to brain atrophy, white matter hyperintensities, and cerebrospinal fluid (CSF) space abnormalities, and compared with neuroradiologist-generated reference reports. Four LLMs (GPT-OSS 20B, MedGemma 4B, Phi-3, and Llama 3.1 8B) generated clinician-facing conclusions from the structured findings. Generated conclusions were evaluated using a five-point Likert scale across four metrics (accuracy, relevance, clarity, and satisfaction). Output errors were assessed across four categories: language, omission, addition, and numeric errors. Structured report findings demonstrated high accuracy for CSF space analysis (77.6%-97.9%). Performance for ordinal visual scales was lower under strict matching (27.1%-57.8%) but improved with a one-grade tolerance (91.7%-97.4%). Among LLMs, GPT achieved the highest scores across most metrics (median score 5) and exhibited the lowest output error rates. Smaller models (Phi-3 and Llama 3.1 8B) showed overall lower performance, with more language, omission, and addition errors. MedGemma demonstrated intermediate performance, but the highest rate of numeric errors. Our quantitative MRI-to-structured reporting pipeline, integrating quantitative MRI software with LLM-based conclusion generation, might be feasible for automated structured drafting of neurodegenerative imaging markers.