Researchers evaluated recent multimodal large language models on the task of medical image alignment assessment, which checks whether two scans show the same anatomy. The study found that GPT‑6 achieved more than 85% accuracy across almost all test scenarios. Older frontier models performed only slightly better than random guessing, while a locally fine‑tuned model matched GPT‑6 on its trained tasks but struggled with new settings. The authors suggest this ability could support quality‑control checks in medical imaging pipelines.
Why it matters
Hospitals could use advanced generalist models to automate image‑quality checks, reducing reliance on scarce specialist operators.
