47 - Comparative analysis of manual and AI assisted cephalometric tracing

Abstract

Introduction

Since its introduction in 1931, cephalometric analysis has played a critical role in orthodontic diagnosis by allowing clinicians to differentiate between skeletal and dentoalveolar patterns of malocclusion, which is essential for appropriate treatment planning. Conventional cephalometric analysis, based on manual landmark tracing, is time-consuming and remains labor-intensive, with digital imaging software still requiring up to 15 minutes for landmark identification. To address this limitation, artificial intelligence (AI) is being integrated into orthodontic software to automate landmark detection. However, the accuracy and reliability of these AI-based systems remain uncertain, and further research is necessary to confirm their accuracy and reliability without human oversight.

Aim

To evaluate techniques for automatic digitization of cephalograms using artificial intelligence (AI) algorithms, highlighting their strengths and weaknesses and reviewing the percentage of success in localizing each cephalometric point.

Methods

Thirty lateral cephalograms from the Mathew’s Collection of the AAOF Legacy Collection were digitized and traced by three calibrated orthodontists. The same radiographs were uploaded to AI-based machine learning programs Orthodx, Angel, and WebCeph. Dolphin software was used to extract x- and y-coordinates for 12 cephalometric landmarks, comprising 8 skeletal and 4 dental points. Mean radial errors (MRE) were assessed relative to thresholds of 1.0 mm, 1.5 mm, and 2 mm to compare successful detection rates (SDR). One-way ANOVA at a significance level of P < .05 was used to compare MRE and SDR. SPSS (IBM v27.0) was used for data analysis.

Results

Experimental results showed that both human tracers and AI-based methods achieved accuracy above 87% using the 2 mm precision threshold, which is considered acceptable in clinical practice. There were no significant differences in MRE or SDR among the evaluated methods based on ANOVA analysis. A marked difference in time was found between the AI-assisted group and the manual group due to heterogeneity in the performance of techniques to detect the same landmark.

Conclusions

Errors in both human tracing and AI-assisted methods showed clinically acceptable accuracy above 87% at the 2 mm threshold. AI assistance may be used in orthodontic practice without compromising accuracy while increasing efficiency in routine clinical and research settings.