Zero-Shot VLM Gemini-3 Reaches 90.3% Accuracy for Condylar CBCT Abnormality Detection

Gemini-3 VLM achieved 90.3% accuracy (95% CI 81.0%-95.5%) for zero-shot detection of condylar osseous changes on CBCT, outperforming GPT-5.2 (75.0%) and Qwen3-VL (55.6%) in 72 internal and 70 external images.

• External validation on a separate dataset (MMDental, n=70) confirmed similar performance, with Gemini-3 accuracy of 90.00%.
• Gemini-3 also generated higher-quality structured reports, outperforming other models across all five QAMAI dimensions (accuracy, clarity, relevance, consistency, and clinical usefulness).
• The models exhibited a conservative diagnostic tendency, suggesting they may miss some subtle abnormalities.

Automated summary

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