Artificial-Intelligence-Based Cephalometric Landmark Detection in Lateral Cephalograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthodontics

Background:

• Accurate landmark identification is essential for reliable cephalometric analysis.
• Traditional manual landmark identification is time-consuming and prone to variability.
• Deep learning offers potential for automating this process.

Purpose of the Study:

• To evaluate a ResNet50-based convolutional neural network for direct automatic localization of 15 landmarks on lateral cephalograms.
• To assess the model's accuracy and reproducibility compared to human annotators.

Main Methods:

• A retrospective study using 669 lateral cephalograms for training and 50 for an internal test set.
• A single-stage regression convolutional neural network (ResNet50) was employed.
• Performance was measured by Euclidean localization errors and success detection rates (SDR).

Main Results:

• The model achieved a mean localization error of 1.25 ± 1.39 mm, with a median error of 0.72 mm.
• Success detection rates within 1.0, 2.0, and 4.0 mm were 64.5%, 81.5%, and 94.4%, respectively.
• Mean error between original and repeated annotations was 0.52 ± 0.92 mm, indicating good inter-observer reliability.

Conclusions:

• The AI model demonstrated promising performance with mean localization error below the 2.0 mm benchmark in an internal test set.
• However, 18.5% of predictions exceeded 2.0 mm, and external validity is unconfirmed.
• Orthodontist verification and external validation are necessary before clinical implementation.