Generative AI for reference-free dentoalveolar aesthetic reconstruction: A multi-center study

Introduction To develop and validate DAR-AI, a clinically applicable artificial intelligence system for reference-free dentoalveolar aesthetic reconstruction of alveolar bone defects. Materials and methods DAR-AI was constructed using a Point Cloud Completion Network integrated with a score-based denoising module to automate and standardize reconstruction according to an anatomical morphology-oriented strategy. The system was validated using data from 170 patients across five centers, with an additional 8 cases used for clinical human-machine comparison. Performance was assessed by geometric accuracy, robustness between internal and external testing sets, blinded human-machine comparison, operation time, and feasibility of integration into clinical workflows. Results and discussion On the external testing set, DAR-AI achieved a Chamfer Distance of 1.191 ± 0.273 mm and a mesh reconstruction RMS of 0.465 ± 0.090 mm, with no statistically significant difference compared with the internal set. In the blinded human-machine comparison, the AI workflow maintained high expert acceptance (most scores > 4.0/5.0) and significantly reduced operation time. The generated 3D models were successfully incorporated into pre-operative communication, 3D-printed bone graft fabrication, and titanium mesh fabrication. Conclusion DAR-AI provides a rapid, accurate, and generalizable solution for dentoalveolar reconstruction by overcoming the constraints of missing anatomical references. This generative AI model demonstrates clinically acceptable planning performance, with high expert acceptance in most assessment dimensions, and holds significant potential for broader applications in complex dentoalveolar aesthetic reconstruction tasks.