Related Experiment Video
Updated: Aug 30, 2026
Three-Dimensional Cephalometric Landmark Annotation Demonstration on Human Cone Beam Computed Tomography Scans
Published on: September 8, 2023
Automated dental landmark detection and model-analysis measurements on 3D digital dental models using a
Myungsoo Bae1, Jae-Woo Park1, Minjung Kim1
1Department of Convergence Medicine and Institute of Digital Healthcare, Asan Medical Center, University of Ulsan College of Medicine, Seoul, Republic of Korea.
Background And Objective:
Accurate localization of dental landmarks in three-dimensional (3D) digital dental models is crucial for orthodontic diagnosis and treatment planning, particularly for model analysis involving measurements such as arch length discrepancy (ALD). Such model analysis is time-consuming and often shows substantial inter-examiner variability. This study proposes a projection-based two-stage cascade convolutional neural network (CNN) for automated landmark detection.
Methods:
We obtained 1427 digital dental models from 714 patients. The primary quantitative evaluation was conducted using a held-out internal test set of 111 complete-dentition models. An external set of 22 models from 11 patients at an independent institution was used for a preliminary external evaluation. Each 3D dental model was converted into five 2D feature maps, including depth, curvature, and the x-, y-, and z-components of the surface normal. A cascade CNN was used, in which the first stage detected tooth regions using RetinaNet and the second stage localized 34 dental landmarks per model using U-Net. The predicted landmarks were mapped back onto the original 3D surface and used to calculate orthodontic model-analysis measurements.
Results:
In the complete-dentition internal test set, tooth-region detection achieved a success rate of 99.44% and a mean IoU of 0.91 ± 0.06. The mean localization error across all predicted landmarks was 0.66 ± 0.49 mm. The mean ALD error calculated from AI-predicted landmarks was 0.90 ± 0.72 mm, comparable to the intra-observer error (0.95 ± 0.78 mm); a higher mean ALD error was observed in the preliminary external evaluation (1.70 ± 0.86 mm). The method also required <4 s per model on average.
Conclusions:
The proposed cascade CNN provides a computationally efficient and accurate approach for automated dental landmark detection and orthodontic model-analysis measurements on 3D digital dental models. The method shows preliminary clinical application potential, but further multicenter, large-scale validation is required before clinical deployment.
