Postoperative Edentulous Patients' Facial Appearance Prediction By Deep Learning Technologies

This study aims to develop a bidirectional deep learning model, FacePointNet, to predict postoperative facial changes in edentulous patients following dental restoration, thereby enhancing preoperative planning and patient satisfaction. Sixteen full-mouth edentulous patients underwent three-dimensional facial scanning preoperatively and six months postoperatively. We developed FacePointNet, a point-set neural network characterized by a dual pointwise alignment sub-network (PR-Nets) within an encoder-decoder architecture, for learning bidirectional geometric transformations between pre- and postoperative facial geometries. The model is optimized using a composite loss function that combines geometric losses (chamfer and point density losses) and transform consistency losses. The performance of the model was evaluated by quadruple cross-validation, where quantitative metrics such as chamfer distance (CD), mean absolute error (MAE) and visual similarity score (VSS) were assessed by three independent oral implantologists to evaluate clinical realism. Translated with www.DeepL.com/Translator (free version) The model had high accuracy in landmark-specific prediction (mean MAE, D3: p = 0.35), whereas there were no significant differences in global geometric similarity (CD for all points, D1: p = 0.14; landmarks, D2: p = 0.63). Expert reviewers reported a high degree of visual consistency (VSS: 4.2/5), especially in key aesthetic areas such as the nasolabial folds. FacePointNet demonstrated good accuracy in modeling local postoperative facial changes, providing clinical utility for preoperative visualization. Future work should expand the dataset, integrate multimodal biomechanical data, and refine dynamic modeling to improve the robustness of global predictions.