The abstract is read at the publisher; the summary is JClub's.
Fine-tuned ModernBERT encoders achieved the highest micro-F1 for classifying 18 chest CT findings in internal Japanese CT-RATE reports, reaching 0.9576 for radiologist-checked text.
- Why it matters: The work addresses how to best classify many CT findings from reports, and when typed yes/no probability decisions can help compared with generative LLMs.
- What they did: Using public retrospective data, the study compared fine-tuned ModernBERT, generative LLMs, and typed decisions on 150 internal CT-RATE reports and 243 external RR-Findings reports, with thresholds fixed on 450 development reports.
- The result: Externally, typed decisions with Jev reached micro-F1 0.8844, exceeded the best generative condition (0.8564) and the Decisions API (0.8177), and showed the smallest internal-to-external drop (-0.0396).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.