Beyond Reporting: Claude 3.7 Sonnet Accurately Classifies T Stage and Uncovers Omitted Anatomic Invasion in Nasopharyngeal Carcinoma Magnetic Resonance Imaging (MRI) Reports

Purpose This study aimed to determine whether Claude 3.7 Sonnet (Anthropic, San Francisco, CA, USA), a large language model (LLM), can (i) assign nasopharyngeal carcinoma (NPC) T classification from routine magnetic resonance imaging (MRI) reports and (ii) identify unreported anatomical structures whose invasion would warrant a higher T stage. Materials and methods This single-institution retrospective study included 38 consecutive patients (31 men; mean age 59.7±14.9 years) who underwent pretreatment MRI for NPC between April 1999 and March 2025. De-identified unstructured "Findings" sections were submitted once to Claude 3.7 Sonnet (temperature=0), prompting the model to assign a T stage according to the American Joint Committee on Cancer/Union for International Cancer Control 9th Edition and to list potentially missed invasive sites. Reference-standard staging and relevant omissions were established independently by two radiologists. Model accuracy for T classification and for detecting missing structures was calculated; false-positive flags were recorded. Radiologists re-evaluated MR images for any stage change prompted by the LLM. Results The LLM reproduced the reference T category in 35/38 patients (92.1%). Category-specific accuracy was 100% for T1 (9/9) and T4 (13/13), 90% for T3 (9/10), and 66.7% for T2 (4/6). Among 208 eligible unmentioned structures, the model correctly flagged 81 (38.9%), with a mean of 3.34 false-positive suggestions per case. Subsequent human review confirmed stage upgrades in 2/38 patients (5.3%), both corrected to T4 based on intracranial extension or cranial nerve involvement noted by the LLM. Conclusion Claude 3.7 Sonnet achieved high accuracy in T staging from unstructured free-text MRI reports for NPC and identified clinically important omissions, enabling radiologists to correct staging in select cases. LLM-assisted report auditing may improve staging quality and serve as an educational aid where subspecialty expertise is limited.