Low certainty of evidence supports the application of (AI) for the automatic detection of cephalometric landmarks with prospects for improvements

Abstract

Study Selection

Electronic search used Embase, IEEE Xplore, LILACS, MedLine (via PubMed), SciELO, Scopus, Web of Science databases, as well as OpenGrey and ProQuest. The search included studies published till November 2021 in any language. Studies written in languages other than English or Portuguese were translated. After removing duplicates, the selection of the studies proceeded by two reviewers independently. Disagreements were resolved with the help of a third reviewer. A reviewer was responsible for the data extraction from the selected studies and a second reviewer did a cross-examination to test the agreement. The risk of individual bias in the eligible studies was assessed independently by two of the authors using QUADAS-2 which includes four domains: patient selection, index test, reference standard, and flow and timing; each of the four domains can be judged as "high risk", “uncertain risk,” or “low risk”. The reviewers resolved the conflict by discussion or by resorting to a third reviewer if the matter is not settled between them.

Key Study Factor

The key study factor was the identification of cephalometric landmarks' from digital images (2D and 3D) by (AI) applications (deep learning and handcrafted) compared to manual identification by experts which is the standard for cephalometric landmarks identification.

Main Outcome Measures

Three main outcome measures were investigated; the agreement (%) of the automatic (AI) and the manual cephalometric landmark identification (2mm and 3mm margin of error) and the divergence (mm) between the identification of the landmarks by the automatic (AI) and the manual methods.