A Dual-Discriminator GAN and Point-Cloud Based Model for 3D Facial Appearance Prediction After Edentulous Implant Rehabilitation

BackgroundLoss of all natural teeth leads to the disappearance of dental and alveolar support, causing collapse of the mid- and lower-facial soft tissues, deepening of nasolabial folds, drooping of the oral commissures, and shortening of the lower facial third. Accurately predicting postoperative facial appearance after implant-supported rehabilitation is critical for aesthetic denture design and patient communication, yet current predictions mainly rely on clinician experience and trial wax-ups. MethodsWe propose a point-cloud-based intelligent method and system for predicting facial appearance after edentulous implant rehabilitation. Three-dimensional facial surface point clouds of edentulous patients are acquired and combined with the 3D morphology of the planned denture. A bidirectional generative adversarial network (GAN) operating on point clouds is constructed. The forward generator GA[-&gt;]B takes the edentulous face A and denture shape C as input and generates the predicted postoperative face [Formula]; the reverse generator GB[-&gt;]A reconstructs the edentulous face A from the postoperative face B, enforcing cycle consistency. The network employs dilated convolutions with skip connections and a dual-discriminator architecture consisting of a global discriminator DG for overall facial morphology and a local discriminator DL focused on the perioral region. The model is trained using a weighted combination of adversarial loss, L1 loss, mean squared error (MSE) loss, and perceptual loss. ResultsThe method was validated using pre- and post-restoration 3D facial scans from 10 edentulous patients treated with implant-supported complete dentures. Qualitative overlays showed close agreement between predicted and real postoperative faces, particularly in lip support, cheek fullness, and chin projection. Quantitatively, most surface points exhibited errors within 1.0 mm. The mean surface distance error was 0.81 mm (standard deviation 0.27 mm), and the root-mean-square error (RMSE) was 0.88 mm. Landmark displacement errors for the nasal tip, upper lip, and chin ranged from 0.5 to 1.0 mm. Predicted increases in lower-facial soft-tissue volume deviated from true values by <8%. Two experienced prosthodontists rated the agreement between predictions and outcomes with a mean visual analogue scale score of 8.5/10. ConclusionThe proposed dual-discriminator point-cloud GAN achieves millimetre-level accuracy in predicting facial soft-tissue changes after edentulous implant rehabilitation without requiring CT imaging or explicit biomechanical modeling. The associated software provides real-time 3D visualization and quantitative analysis, offering an effective preoperative tool to optimize denture design and enhance patient satisfaction.