Predicting 3D Post-Orthodontic Facial Outcomes With a Diffusion Model Trained on Unpaired Datasets.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Orthodontics

Background:

• Predicting post-orthodontic facial aesthetics is vital for clinical planning.
• Traditional methods lack accuracy and realism in facial morphology prediction.
• A novel generative AI framework using a diffusion model is introduced for patient-specific 3D facial morphology prediction.

Purpose of the Study:

• To develop and validate a generative AI framework for predicting 3D facial morphology after orthodontic treatment.
• To assess the accuracy and perceptual realism of the AI model using unpaired pre- and post-treatment datasets.

Main Methods:

• A denoising diffusion implicit model (DDIM) was trained on 238 pre-treatment and 245 post-treatment cone-beam computed tomography (CBCT) scans.
• Model validation used 30 paired CBCT scans, evaluating Euclidean distance errors at 13 landmarks and lateral profile metrics.
• Perceptual realism was assessed via a visual Turing test with orthodontists.

Main Results:

• The model achieved a mean Euclidean error of 1.22 ± 0.75 mm at soft tissue landmarks.
• 91.03% of predictions were within the clinically acceptable 2 mm threshold.
• Orthodontists could not reliably distinguish AI-predicted outcomes from actual results in a visual Turing test.

Conclusions:

• The diffusion-based generative model accurately predicts post-orthodontic 3D facial changes from unpaired data.
• This AI framework shows potential as an auxiliary tool for visualizing treatment outcomes.
• The model can improve patient-clinician communication and manage treatment expectations.