[A preliminary exploration of an intelligent system for personalized tooth morphology reconstruction based on deep learning].

Objective: To integrate implicit templates with deep learning techniques, a novel neural network, the tooth-deformable deep implicit network (T-DDIN), was constructed to achieve high-precision shape completion of tooth defects in a personalized manner. Methods: A total of 550 intraoral scan models were collected from patients treated at the Department of Orthodontics and Department of Prosthodontics, West China Hospital of Stomatology, Sichuan University (500 for training and 50 for testing), between March 2022 and March 2024. T-DDIN reconstructed defective tooth morphology using an implicit template and a latent encoding prediction network. During model evaluation, Class Ⅱ cavity defects and occlusal wear defects were simulated in the test set. Morphological restoration was performed using both traditional computer aided design (CAD) methods and the T-DDIN deep learning approach. The two methods were compared based on three-dimensional deviation, occlusal adjustment volumes, cusp angle deviation, and restoration time. Results: The T-DDIN group demonstrated significantly lower three-dimensional deviation for Class Ⅱ cavity defects and occlusal wear restoration [(0.14±0.05) and (0.16±0.09) mm], occlusal adjustment volumes [(0.44±0.03) and (0.49±0.03) mm³], and difference value of the tooth cusp angles (5.69°±1.90° and 6.04°±0.53°) compared to the traditional CAD group (both P<0.001). No significant differences were observed within the T-DDIN group between the two defect types in terms of three-dimensional deviation (P=0.098) or occlusal adjustment volume (P=0.154) or difference value of the tooth cusp angles (P=0.196). However, in the traditional CAD group, three-dimensional deviation, occlusal adjustment volume and difference value of the tooth cusp angles was significantly higher in occlusal wear restorations than in Class Ⅱ cavity defects restorations (P<0.001). The T-DDIN group, which involved Class Ⅱ cavity defects and occlusal wear, demonstrated significantly less recovery time of morphology (37.2±7.7) and (39.4±6.2) s compared to the traditional CAD group (P<0.001). Conclusions: T-DDIN demonstrated superior stability and accuracy in morphological reconstruction for various types of dental defects while significantly reducing restoration time.