Skel-Net: automatic prediction of skeletal pattern on scanned lateral cephalograms using anatomical prior-guided deep...

Area of Science:

• Craniofacial development research
• Artificial intelligence in orthodontics
• Medical image analysis

Background:

• Accurate craniofacial pattern estimation is vital for orthodontic success.
• Static measurements and manual cephalometric analysis have limitations in capturing dynamic changes and require expertise.
• Skel-Net is introduced as a novel deep learning network for estimating craniofacial growth patterns.

Purpose of the Study:

• To develop and validate Skel-Net, an anatomic prior-guided deep learning network.
• To estimate changes in the ANB angle over five years in pediatric patients (aged 8-16).
• To improve the accuracy and efficiency of craniofacial growth prediction for orthodontic applications.

Main Methods:

• A two-stage approach combining cephalometric landmark detection (Ceph-Net) with multichannel inputs.
• Utilized two-dimensional heatmaps and ANB angle priors to enhance prediction accuracy.
• Trained and validated on 612 lateral cephalograms from 245 patients, comparing against established deep learning models.

Main Results:

• Skel-Net demonstrated superior performance compared to DenseNet121, MobileNetV2, ResNet101, and VGG16.
• Achieved the lowest prediction errors: Mean Absolute Error of 1.021 degrees and Root Mean Squared Error of 1.338 degrees.
• Recorded the highest R-squared value (0.517), indicating robust predictive capabilities for craniofacial growth.

Conclusions:

• Skel-Net effectively leverages anatomic priors and longitudinal data for dynamic, personalized craniofacial growth predictions.
• The framework supports early and precise orthodontic interventions, enhancing treatment efficiency and stability.
• This AI-driven approach promises to improve overall patient outcomes in orthodontics.