Experimental analysis of the accuracy of cephalometric landmark identification in lateral teleroentgenograms

Objective. To evaluate the promising application of neural networks for cephalometric analysis by analyzing the accuracy of manual identification of anatomical landmarks on digital lateral teleradiographs. Materials and Methods. Markup of 100 anonymized teleradiographs in lateral projection by eleven orthodontists on 21 parameters was performed, 23100 digital X-ray images with a reference point plotted on them were obtained. The coordinates of the reference point were compared with the “base point”, i.e. the averaged coordinate for each reference point among all its localizations. Results. According to the criterion of average deviation from the “base point”, the best accuracy was achieved for the apices of the incisal edges of the central incisors of the maxilla (is) (0.589, CI = 95%) and mandible (ii) (0.835, CI = 95%), as well as for the middle of the entrance to the Turkish saddle (S) (0.662, CI = 95%). For the group of landmarks with the lowest consistency, which included points such as Po (4.330, CI = 95%), Pt (2.999, CI = 95%) and Ba (2.887, CI = 95%), the use of artificial neural networks alone is likely to be insufficient to automate identifications and improve the quality of cephalometric analysis and other machine learning elements will need to be implemented. Conclusion. Considering the results of our study, we can conclude that the proposed method demonstrates high accuracy for most points and can be used to automate cephalometric analysis with further development of machine learning technologies.