6

Introduction : As well as being a rather tedious task, manual cephalometric analyses are prone to reproducibility issues (tracing and measurement errors), making the prospect of fully automated digital tracing techniques highly attractive. The objectives of the study were as follows: to evaluate the positioning (accuracy and precision) of common cephalometric landmarks by two software programs offering artificial intelligence (AI) detection (WebCeph™ and DentaliQ®) compared to a manual reference; and then to the compare the two software programs. Materials and method: Sixty-eight lateral cephalograms were selected for this study. Twenty-two landmarks were manually annotated, and then the discrepancies between this gold standard and the points detected by each AI software were measured, along with the success detection rate (SDR). Statistical analyses were carried out using “confidence ellipses” and two-tailed t-test (p-value of 5%). Results : In terms of precision, the average SDRs within 2 mm for WebCeph™ and DentaliQ® were 57.2% and 66.5%, respectively. In terms of accuracy, the best results were obtained for S, Na, and the incisal edges. Significant random errors were found for the points Po, So, ENA, ENP, Ba, and Go. Other points, such as Pog and B, exhibited considerable vertical dispersion. Overall, DentaliQ® had a slight advantage, although the difference was not significant. Discussion: The measured detection accuracy appears insufficient for unsupervised use. Indeed, the results are promising for the detection of certain points, and AI could help clinicians save time, but the software should never have the final say, and the ability to reposition points must remain possible. Conclusion: This technology is advancing rapidly, and effective clinical use will probably be possible in the near future.