Accuracy of cephalometric landmark identification by artificial intelligence platform versus expert orthodontist in...

The primary aim of the study was to evaluate the accuracy of automated artificial intelligence (AI) cephalometric landmark identification in cleft patients and compare it to landmarks identified by an expert orthodontist. The secondary objective was to compare cephalometric measurements obtained by both methods.

A total of 112 pre-treatment lateral cephalometric radiographs of unilateral cleft palate patients were collected from the archives of the Department of Orthodontics, Faculty of Dentistry, Alexandria University following screening of all the records of patients treated in the period January 2019-December 2022 for eligibility. For each of the acquired radiographs, cephalometric tracing was performed by fully automated WebCeph™ landmark detection process and by manual identification of the landmarks by an expert orthodontist using OnyxCeph™ software. The traced radiographs were then imported into Photoshop software for evaluation of the (x,y) coordinates, in mm, for each of the identified landmarks (Primary outcome). Moreover, linear and angular measurements generated using WebCeph™ and OnyxCeph™ software were compared (secondary outcomes).

The coordinates of A-point, ANS, and Or showed statistically significant differences between both identification methods, with a mean difference between the two methods ranging between -0.86mm±2.15 and 3.15mm±6.07. None of the dental landmarks showed statistically significant differences between the two methods. None of the soft tissue landmarks showed statistically significant differences, except Ns y-coordinate. Several points showed clinically significant differences between both methods. The greatest mean differences in cephalometric measurements between the two methods were reported in nasolabial angle CotgSnLs (18.3±22.77̊) followed by Max1-NA (-8.86±17.46̊) and Max1-SN (-8.43±12.51̊).

The identification of cephalometric landmarks in cleft palate patients using the web-based AI platform is not as accurate as manual identification. Manual adjustment of landmarks following AI identification is advised.