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

Background: Accurate prediction of facial aesthetics after orthodontic treatment is crucial for clinical planning, yet traditional methodologies often lack the required accuracy and realism. We introduce a novel generative artificial intelligence framework utilizing a diffusion model to predict patient-specific 3D facial morphology, specifically designed to function with unpaired pre- and post-treatment datasets. Methods: This retrospective study utilized non-paired pre-treatment (n = 238) and post-treatment (n = 245) cone-beam computed tomography (CBCT) scans for model training. A discrete test set, comprising 30 paired pre- and post-treatment CBCT scans, was employed for validation. We developed a denoising diffusion implicit model (DDIM) engineered to learn the transformation from a pre-treatment 3D facial mesh to a predicted post-treatment outcome. The model's predictive accuracy was quantitatively evaluated through Euclidean distance errors at 13 soft tissue landmarks, analysis of lateral profile metrics, and measurement of the mean surface distance. Perceptual realism was assessed via a visual Turing test administered to 3 experienced orthodontists. Results: The model demonstrated high predictive accuracy, yielding a mean Euclidean error of 1.22 ± 0.75 mm across all evaluated landmarks. The successful prediction rate within the clinically acceptable 2 mm threshold was 91.03%. No statistically significant differences were observed between the predicted and actual outcomes for seven key lateral profile measurements. In the visual Turing test, the mean identification accuracy of the orthodontists was 52.22%, a result approximating random chance. Conclusion: The proposed diffusion-based model is capable of generating accurate and perceptually realistic 3D predictions of post-orthodontic facial changes, even when trained on unpaired datasets. Clinical Significance: The results suggest that this generative framework holds potential as an auxiliary tool for visualizing post-orthodontic facial changes. By facilitating patient-clinician communication and helping to manage treatment expectations, the model offers a valuable, data-driven reference to complement professional clinical judgment.