Related Experiment Video
Updated: May 10, 2026
Three-Dimensional Cephalometric Landmark Annotation Demonstration on Human Cone Beam Computed Tomography Scans
Published on: September 8, 2023
A new automated 3d facial soft tissue landmarking method via deep learning
Han Bao1, Zhidong He2, Jiasong Wu3
1Department of Orthodontics, The Affiliated Stomatological Hospital of Nanjing Medical University, Nanjing 210029, China; State Key Laboratory Cultivation Base of Research, Prevention and Treatment for Oral Diseases (Nanjing Medical University), Nanjing 210029, China; Jiangsu Province Engineering Research Center of Stomatological Translational Medicine (Nanjing Medical University), Nanjing 210029, China.
Objectives:
To propose and evaluate a novel deep learning (DL)-based facial soft tissue network (FST-Net) for the automatic detection of 34 commonly used three-dimensional (3D) soft tissue landmarks and eight measurements.
Methods:
A total of 297 patient cases were randomly allocated into training (n = 207), validation (n = 30), and testing sets (n = 60) for automated landmark detection through feature fusion and local coordinate regression. The model's performance in terms of landmark localisation and soft tissue measurements was evaluated.
Results:
The mean radial error (MRE) of all 34 landmarks was 1.80 ± 1.81 mm. For the 28 landmarks with high operator reproducibility (Group 1 landmarks), the MRE was 1.25 ± 0.90 mm, whereas the six landmarks dependent on subjective judgement (Group 2 landmarks) had a larger average error (4.39 ± 2.57 mm). The model achieved successful detection rates (SDRs) within 2, 2.5, 3, and 4 mm of 72.16%, 80.69%, 85.10%, and 90.54%, respectively. By sub-region, lip landmarks had the highest accuracy (1.03 ± 0.69 mm), whereas bilateral cheek landmarks had the lowest (4.39 ± 2.57 mm). The average intra-class correlation coefficient (ICC) values for eight soft tissue measurements were 0.909 between manual and automated detection. Six out of the eight measurements had average errors below 2 mm or 2° CONCLUSION: These findings underscore the potential of deep learning in orthodontics, as exemplified by FST-Net, which offers significant improvements in the accurate and efficient detection of 3D soft tissue landmarks. The method for Group 1 landmarks achieved satisfactory results, whereas for Group 2 landmarks, further optimisation is needed to improve accuracy in future studies.
Clinical Significance:
FST-Net can automate the detection of 34 commonly used 3D soft tissue landmarks to measure 3D facial characteristics accurately and efficiently, for diagnosis and treatment planning through feature fusion and local coordinate regression. It provides a rapid preliminary landmarking for distinct anatomical features, serving as an auxiliary tool to reduce manual workload, though clinician verification remains mandatory for lateral facial regions.
