LLM-Driven Extraction of NI-RADS and Imaging Tumor Characteristics to Enhance Oropharyngeal Cancer Survivorship Surveillance

The study examined the use of a large language model (LLM) based on GPT-5 to extract information from unstructured radiological reports of oropharyngeal cancer patients. The research team analyzed 200 reports from a total of 7,076 eligible examinations using CT, MRI, or PET/CT scans. The LLM was tested for extracting NI-RADS (Neck Imaging Reporting and Data System) scores and key disease characteristics. In detecting absence of disease, the model achieved 93.3% agreement for primary tumor and 90.3% for lymph nodes. In detecting disease presence, agreement was 94.9% for primary tumor, 89.1% for nodal involvement, and 94.7% for distant metastases. Specificity was high (0.95-0.99) with negative predictive value of 0.99 for distant metastases. Results demonstrate that the LLM-based approach is suitable for standardized and scalable surveillance monitoring of head and neck cancer patients.