LLMs tested on 93 mixed dentition radiographs for tooth ID, caries, and periapical assessment

Large language models evaluated on 93 mixed dentition radiographs from children: Gemini 3 Pro had highest accuracy for tooth identification and caries detection, but none could identify periodontitis. All LLMs were inferior to expert human judgment.

• Gemini 3 Pro outperformed GPT-5 and Sonar for tooth detection/identification and caries level agreement.
• No LLM could identify periodontitis; GPT-5 and Sonar were better at identifying sound teeth, while Gemini 3 Pro recognized advanced carious lesions.
• Overall, all LLMs were inferior to expert pediatric dentists' judgment.

Automated summary

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