Related Experiment Video
Updated: May 28, 2026
10:23
Three-Dimensional Cephalometric Landmark Annotation Demonstration on Human Cone Beam Computed Tomography Scans
Published on: September 8, 2023
Clinical Equivalence of a CNN-Based Automated Soft Tissue Landmark Detection System on 2D Facial Images
Argun Ege Türkün1, Müslim Ege Kalender2, Murat Kurt2
1Department of Orthodontics, Faculty of Dentistry, Ege University, 35040 Izmir, Turkey.
Diagnostics (Basel, Switzerland)
|May 27, 2026
Summary
A new deep learning model using convolutional neural networks (CNNs) shows high accuracy in identifying soft tissue landmarks on 2D facial images for orthodontics. This AI approach significantly speeds up landmark identification and improves reliability compared to manual methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Orthodontics
Background:
- Accurate identification of soft tissue landmarks on 2D facial images is crucial for orthodontic diagnosis and treatment planning.
- Manual landmark identification is time-consuming and can be subject to inter-observer variability.
- Deep learning models offer potential for automating and improving the efficiency of facial landmark detection.
Purpose of the Study:
- To compare the accuracy, reliability, and time efficiency of a CNN-based deep learning model against manual annotation for soft tissue landmark identification in orthodontics.
- To evaluate the performance of the developed CNN model on 2D facial images extracted from 3D scans.
- To assess the potential of AI in streamlining orthodontic digital workflows.
Main Methods:
- Acquired 3D facial scans from 100 participants (18-25 years old).
- Extracted frontal and profile 2D images from 3D models.
- Manually annotated 22 frontal and 15 profile landmarks using LabelMe software.
- Developed and trained a novel CNN model on the manually annotated images.
- Compared CNN model's automatic landmark identification with manual annotations for positional error, time, and reproducibility.
Main Results:
- The CNN model achieved a mean localization accuracy of 96.07% with prediction errors ranging from 2.3% to 4.5%.
- Certain landmarks (Trichion, Menton, Gonion) showed higher error rates.
- The CNN model significantly reduced annotation time (237s per image for manual method).
- Manual landmarking showed excellent intra-observer reliability (ICC: 0.85-0.95).
- The AI model demonstrated consistent predictions.
Conclusions:
- The CNN-based deep learning model offers comparable accuracy to manual landmark identification in orthodontics.
- The AI model significantly enhances annotation speed and reproducibility.
- CNN-based systems present a promising alternative for clinical orthodontic analysis and digital workflow integration.
