GPT-5-Thinking Matches Subspecialist Accuracy, Lifts Generalist Performance in Orbital/Head-Neck Tumor MRI

Generative Pretrained Transformer (GPT)-5-Thinking matched subspecialist top-1 accuracy in diagnosing orbital (76.7% vs 78.3%, P=.64) and head-and-neck tumors (67.5% vs 68.0%, P=.92). Generalist accuracy rose from 61.4% to 70.3% for orbital and 47.0% to 61.0% for head-and-neck t…

• Dual-center retrospective study of 1000 patients with pathologically confirmed orbital or head-and-neck tumors and pretreatment MRI reports.
• Phase 1: GPT-5-Thinking top-1 accuracy was 76.7% (orbital) and 67.5% (head/neck), similar to subspecialists (P=.64 and P=.92).
• Phase 2: With GPT-5-Thinking, generalists’ accuracy rose from 61.4% to 70.3% (orbital) and 47.0% to 61.0% (head/neck) (P<.001).

Automated summary

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