Abstract：
The transition from manual to automatic cephalometric landmark identification has not yet reached a consensus for clinical application in orthodontic diagnosis. The present umbrella review aimed to assess artificial intelligence (AI) performance in automatic 2D and 3D cephalometric landmark identification.
A combination of free text words and MeSH keywords pooled by boolean operators: Automa* AND cephalo* AND (\"artificial intelligence\" OR \"machine learning\" OR \"deep learning\" OR \"learning\").
A search strategy without a timeframe setting was conducted on PubMed, Scopus, Web of Science, Cochrane Library and LILACS.
The study protocol followed the PRISMA guidelines and the PICO question was formulated according to the aim of the article. The database search led to the selection of 15 articles that were assessed for eligibility in full-text. Finally, 11 systematic reviews met the inclusion criteria and were analyzed according to the risk of bias in systematic reviews (ROBIS) tool.
AI was not able to identify the various cephalometric landmarks with the same accuracy. Since most of the included studies\' conclusions were based on a wrong 2 mm cut-off difference between the AI automatic landmark location and that allocated by human operators, future research should focus on refining the most powerful architectures to improve the clinical relevance of AI-driven automatic cephalometric analysis.
Despite a progressively improved performance, AI has exceeded the recommended magnitude of error for most cephalometric landmarks. Moreover, AI automatic landmarking on 3D CBCT appeared to be less accurate compared to that on 2D X-rays. To date, AI-driven cephalometric landmarking still requires the final supervision of an experienced orthodontist.
A combination of free text words and MeSH keywords pooled by boolean operators: Automa* AND cephalo* AND (\"artificial intelligence\" OR \"machine learning\" OR \"deep learning\" OR \"learning\").
A search strategy without a timeframe setting was conducted on PubMed, Scopus, Web of Science, Cochrane Library and LILACS.
The study protocol followed the PRISMA guidelines and the PICO question was formulated according to the aim of the article. The database search led to the selection of 15 articles that were assessed for eligibility in full-text. Finally, 11 systematic reviews met the inclusion criteria and were analyzed according to the risk of bias in systematic reviews (ROBIS) tool.
AI was not able to identify the various cephalometric landmarks with the same accuracy. Since most of the included studies\' conclusions were based on a wrong 2 mm cut-off difference between the AI automatic landmark location and that allocated by human operators, future research should focus on refining the most powerful architectures to improve the clinical relevance of AI-driven automatic cephalometric analysis.
Despite a progressively improved performance, AI has exceeded the recommended magnitude of error for most cephalometric landmarks. Moreover, AI automatic landmarking on 3D CBCT appeared to be less accurate compared to that on 2D X-rays. To date, AI-driven cephalometric landmarking still requires the final supervision of an experienced orthodontist.
摘要：
目的：从手动到自动头影标志识别的转变尚未达成临床应用于正畸诊断的共识。本综述旨在评估自动2D和3D头影测量界标识别中的人工智能（AI）性能。
方法：由布尔运算符汇集的自由文本单词和MeSH关键字的组合：Automa*ANDcaltho*AND（\"人工智能\"或\"机器学习\"或\"深度学习\"或\"学习\"）。
方法：在PubMed上进行了没有时间范围设置的搜索策略，Scopus,WebofScience,Cochrane图书馆和LILACS。
方法：研究方案遵循PRISMA指南，并根据文章的目的制定了PICO问题。数据库搜索导致选择了15篇文章，这些文章被评估为全文的资格。最后,11篇系统评价符合纳入标准，并根据系统评价（ROBIS）工具中的偏倚风险进行分析。
结论：AI无法以相同的准确性识别各种头颅标志。由于大多数纳入的研究结论都是基于AI自动地标位置与人工操作员分配的位置之间错误的2毫米截止差异，未来的研究应该集中在完善最强大的架构，以提高AI驱动的自动头颅测量分析的临床相关性。
结论：尽管性能逐步提高，对于大多数头影测量标志，AI已超过建议的误差幅度。此外，与2DX射线相比，3DCBCT上的AI自动标记似乎不太准确。迄今为止,AI驱动的头影测量标记仍然需要经验丰富的正畸医生的最终监督。
方法：由布尔运算符汇集的自由文本单词和MeSH关键字的组合：Automa*ANDcaltho*AND（\"人工智能\"或\"机器学习\"或\"深度学习\"或\"学习\"）。
方法：在PubMed上进行了没有时间范围设置的搜索策略，Scopus,WebofScience,Cochrane图书馆和LILACS。
方法：研究方案遵循PRISMA指南，并根据文章的目的制定了PICO问题。数据库搜索导致选择了15篇文章，这些文章被评估为全文的资格。最后,11篇系统评价符合纳入标准，并根据系统评价（ROBIS）工具中的偏倚风险进行分析。
结论：AI无法以相同的准确性识别各种头颅标志。由于大多数纳入的研究结论都是基于AI自动地标位置与人工操作员分配的位置之间错误的2毫米截止差异，未来的研究应该集中在完善最强大的架构，以提高AI驱动的自动头颅测量分析的临床相关性。
结论：尽管性能逐步提高，对于大多数头影测量标志，AI已超过建议的误差幅度。此外，与2DX射线相比，3DCBCT上的AI自动标记似乎不太准确。迄今为止,AI驱动的头影测量标记仍然需要经验丰富的正畸医生的最终监督。
