人工智能辅助根尖片评估:病变检测、牙髓并发症分析与临床治疗建议综述
该回顾性诊断准确性研究比较了3种视觉语言AI模型(ChatGPT-5 Plus、Gemini 2.5 Pro、Copilot Pro)与牙髓病专家在根尖片诊断中的表现,涵盖治疗选择、并发症识别和病变检测。Gemini 2.5 Pro诊断准确性最高,尤其在根尖周病变检测中灵敏度100%、特异度88%;Copilot Pro对穿孔和器械折断等并发症灵敏度明显较低。AI在治疗选择和病变检测方面表现较好,但在复杂操作并发症识别上可靠性不足。
要点
- Gemini 2.5 Pro在根尖周病变检测中灵敏度100%、特异度88%,表现最均衡。
- Copilot Pro对穿孔和器械折断等并发症灵敏度明显较低。
- AI在治疗选择任务中表现较好,但复杂并发症识别可靠性不足。
核心结论
- AI系统在治疗选择和病变检测任务中展现出较好的诊断准确性,但在识别复杂操作并发症方面可靠性较低。
- Gemini 2.5 Pro表现最均衡,Copilot Pro在不同诊断类别间变异性最高。
研究设计与样本
- 回顾性诊断准确性研究,评估根尖片与治疗决策、操作并发症及病变检测的关系。
- 以牙髓病专家作为参考金标准,比较3种视觉语言AI模型(ChatGPT-5 Plus、Gemini 2.5 Pro、Copilot Pro)与专家的诊断表现。
- 诊断类别包括初治选择、非手术再治疗、最终治疗决策、穿孔、欠充、超充、器械折断、钙化和根尖周病变检测。
主要发现
- 牙髓病专家之间的一致性几乎完美(κ = 0.95)。
- Gemini 2.5 Pro诊断准确性最高,尤其在根尖周病变检测中灵敏度100%、特异度88%。
- ChatGPT-5 Plus在治疗选择方面表现同样较强。
- Copilot Pro对穿孔和器械折断等并发症的灵敏度明显较低。
- Gemini和ChatGPT-5 Plus在术前和术后治疗决策的Kappa值较高,而Copilot Pro较低;Friedman检验证实组间差异显著(P < .001)。
关键数值与推荐
- 牙髓病专家间一致性κ = 0.95。
- Gemini 2.5 Pro根尖周病变检测灵敏度100%、特异度88%。
- Friedman检验组间差异P < .001。
临床怎么用
- AI可辅助治疗选择和根尖周病变检测,但复杂并发症识别仍需专家把关。
- 不同AI模型诊断可靠性差异较大,临床选用需关注其验证表现。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
INTRODUCTION: Artificial intelligence (AI) systems are increasingly used in dental radiology to support endodontic diagnosis. However, their diagnostic reliability across different clinical categories remains unclear. This study compared 3 vision-language AI models (ChatGPT-5 Plus, Gemini 2.5 Pro, and Copilot Pro) with expert endodontists by assessing sensitivity, specificity, overall diagnostic agreement, and Youden's Index across multiple endodontic conditions. METHODS: This retrospective diagnostic accuracy study evaluated the relationship between periapical radiographs and treatment decisions, procedural complications, and lesion detection. Expert endodontists served as the gold standard of reference. Diagnostic categories included primary treatment selection, nonsurgical retreatment, final treatment decisions, perforation, underfilling, overfilling, broken file, calcification, and periapical lesion detection. RESULTS: There was an almost perfect agreement between the endodontists (κ = 0.95). Gemini 2.5 Pro demonstrated the highest diagnostic accuracy, particularly in periapical lesion detection (sensitivity 100%, specificity 88%), while ChatGPT-5 Plus showed similarly strong performance in treatment selection. Copilot Pro exhibited markedly low sensitivity for complications such as perforation and instrument fracture. Kappa values for preoperative and postoperative treatment decisions were high for Gemini and ChatGPT-5 Plus, but low for Copilot Pro. The Friedman test confirmed significant differences among the groups (P < .001). CONCLUSIONS: AI systems demonstrated promising diagnostic accuracy in treatment selection tasks and lesion detection, but performed less reliably in identifying complex procedural complications. Gemini 2.5 Pro showed the most balanced performance, whereas Copilot Pro displayed the highest variability across diagnostic categories.
引用
- Akyüz İE, Kıvırcık BE, Aslan T, et al. Artificial Intelligence-assisted Periapical Radiographic Assessment: Lesion Detection, Endodontic Complication Analysis, and Review of Clinical Treatment Recommendations. J Endod. 2026. PMID: 42019676. doi:10.1016/j.joen.2026.04.008
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。