Feasibility of predicting vertical cephalometric angles from panoramic radiographs using deep learning.

Area of Science:

• Orthodontics
• Radiology
• Artificial Intelligence

Background:

• Early identification of vertical skeletal discrepancies is crucial for orthodontic diagnosis and treatment planning.
• Panoramic radiographs (OPGs) are more common than lateral cephalometric radiographs (LCRs).

Purpose of the Study:

• To evaluate the potential of artificial intelligence (AI) to predict vertical skeletal angles using OPGs.
• To compare the performance of various deep learning models in estimating these angles.

Main Methods:

• Retrospective collection of 715 patient LCRs and OPGs.
• Training multiple convolutional neural network (CNN) architectures (EfficientNet-B3, DenseNet, ResNet, VGG) and ensemble averaging to predict Frankfort-mandibular plane angle (FMA), gonial angle, and Sum of Björk from OPGs.
• Performance evaluation using Mean Absolute Error (MAE), coefficient of determination (R²), and intraclass correlation coefficients (ICC).

Main Results:

• Ensemble averaging demonstrated the highest predictive accuracy with MAE values of 2.53°±0.08 for FMA, 3.16°±0.10 for gonial angle, and 3.04°±0.09 for Sum of Björk.
• High agreement was observed among different CNN architectures (P<0.001).
• Gradient-weighted Class Activation Mapping (Grad-CAM) indicated that predictions focused on the gonial angle region, condylar area, and mandibular ramus.

Conclusions:

• Deep learning shows potential for estimating vertical skeletal measurements from OPGs.
• Current prediction errors indicate that AI cannot yet replace cephalometric analysis.
• Future improvements may involve larger datasets, geometry-aware models, and external validation to enhance predictive accuracy.