Gemini 3.1 Pro outperforms GPT-5.1 at Cobb angle measurement, but both lack accuracy for autonomous use

In a study of 125 spine radiographs, Gemini 3.1 Pro measured Cobb angles with lower error than GPT-5.1 (MAE 5.06° vs 9.62°, P<0.001), but agreement was still insufficient for autonomous use; expert verification remains necessary.

• Gemini 3.1 Pro achieved ICC 0.842 vs. GPT-5.1's 0.587, and 61.6% of Gemini measurements were within 5° of reference vs. 24.8% for GPT.
• Categorical severity agreement was 80.0% for Gemini vs. 52.8% for GPT, but end-vertebra concordance was <20% for both models.
• The study used a single-session protocol, and neither model's repeatability was assessed.

Automated summary

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