Privacy-Preserving LLM Achieves 99.8% Detection of Extraspinal Findings on MRI Spine Reports

A privacy-preserving large language model detected 99.8% of extraspinal findings on MRI spine reports, with 96.3% accuracy in referral decisions and near-perfect agreement (κ=0.959) with the reference standard.

• Out of 400 MRI spine reports (395 patients), 58.7% had at least one extraspinal finding; renal and gynecological findings were most common.
• The model correctly identified all 12 clinically urgent findings (e.g., aortic dissection).
• Misclassification: 2% of findings downgraded to lower C-RADS categories, 1% upgraded to higher categories. Analysis completed in under 5 seconds per report.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.