GPT-5 Extracts NI-RADS and Disease Status from Oropharyngeal Cancer Surveillance Reports with High Negative Agreement

An LLM (GPT-5) extracted pNI-RADS and nNI-RADS scores from OPC surveillance reports with 93.3% and 90.3% agreement for NI-RADS 1 (no disease). For NI-RADS ≥2, exact category agreement was 73.1% (pNI-RADS) and 64.3% (nNI-RADS). Disease-status specificity was 0.95–0.99. Retrospect…

• Retrospective study of 200 post-treatment head/neck imaging reports, with three-reader consensus as the reference standard.
• For disease-status variables, agreement was 94.9% (primary tumor, F1 0.87), 89.1% (nodal disease, F1 0.87), and 94.7% (distant metastasis, F1 0.70); specificity ranged 0.95–0.99.
• Limitations: single-center, no external validation, and evaluation performed on the same dataset used to develop the extraction pipeline.

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