Automatic Extraction of Structured Information from Brain MRI Reports Using an Open-Weight Large Language Model

In three linesStudy on automatic extraction of structured data from Dutch brain MRI reports. LLaMA 3.1 achieves 90-96% accuracy for visual rating scores (medial temporal atrophy, global cortical atrophy, Fazekas), 93% for microbleed detection, but 66-80% for numerical variables. Few-shot prompting improves numerical performance (92% for microbleeds).

Summary generated by Claude — human-verified