Deep Learning-Based Prediction of the 3D Postorthodontic Facial Changes

Area of Science:

• Orthodontics and Dental Imaging
• Artificial Intelligence in Healthcare
• 3D Facial Reconstruction

Background:

• Increasing adult orthodontic population necessitates accurate 3D posttreatment facial prediction.
• Current prediction methods lack precision and evidence-based validation.
• Need for integrating patient-specific factors and treatment conditions into predictive models.

Purpose of the Study:

• To develop a novel 3D postorthodontic face prediction method using deep learning.
• To validate the accuracy and clinical applicability of the proposed AI-driven prediction.
• To leverage cone-beam computed tomography (CBCT) data for enhanced facial outcome prediction.

Main Methods:

• A conditional generative adversarial network (cGAN) was trained on 268 paired pretreatment (T1) and posttreatment (T2) CBCT scans.
• Input variables included patient gender, age, and changes in upper (ΔU1) and lower incisor positions (ΔL1).
• Accuracy was assessed using prediction error, mean absolute distances at perioral landmarks, and percentage of error < 2 mm on a test set (n=44).

Main Results:

• The predicted posttreatment (PT2) 3D faces closely resembled actual posttreatment (T2) faces, particularly in perioral regions.
• The mean prediction error was 1.2 ± 1.01 mm, with 80.8% of predictions achieving accuracy within acceptable limits.
• Over 50% of experienced orthodontists could not differentiate between real and AI-predicted posttreatment facial images.

Conclusions:

• A valid and accurate 3D postorthodontic face prediction method was successfully developed using deep learning.
• The AI model demonstrates significant potential for clinical usability in orthodontic treatment planning.
• CBCT data combined with deep learning offers a powerful approach for predicting orthodontic treatment outcomes.