Facial Surgery Preview Based on the Orthognathic Treatment Prediction

作     者：Han, Huijun Zhang, Congyi Zhu, Lifeng Singh, Pradeep Hsung, Richard Tai-Chiu Leung, Yiu Yan Komura, Taku Wang, Wenping Gu, Min 

作者机构：Discipline of Orthodontics Faculty of Dentistry the University of Hong Kong Hong Kong Department of Computer Science Faculty of Engineering the University of Hong Kong Hong Kong School of Instrument Science and Engineering Southeast University Nanjing China Discipline of Oral and Maxillofacial Surgery the University of Hong Kong Hong Kong Department of Computer Science and Engineering Texas A&M University Texas United States 

出 版 物：《arXiv》 (arXiv)

年 卷 期：2024年

核心收录：

主　　题：Transplantation (surgical) 

摘      要：Background and Objective: Orthognathic surgery consultations are essential for helping patients understand how their facial appearance may change after surgery. However, current visualization methods are often inefficient and inaccurate due to limited pre- and post-treatment data and the complexity of the treatment. This study aims to develop a fully automated pipeline for generating accurate and efficient 3D previews of postsurgical facial appearances without requiring additional medical images. Methods: The proposed method incorporates novel aesthetic criteria, such as mouth-convexity and asymmetry, to improve prediction accuracy. To address data limitations, a robust data augmentation scheme is implemented. Performance is evaluated against state-of-the-art methods using Chamfer distance and Hausdorff distance metrics. Additionally, a user study involving medical professionals and engineers was conducted to evaluate the effectiveness of the predicted models. Participants performed blinded comparisons of machine learning-generated faces and real surgical outcomes, with McNemar’s test used to analyze the robustness of their differentiation. Results: Quantitative evaluations showed high prediction accuracy for our method, with a Hausdorff Distance of 9.00 millimeters and Chamfer Distance of 2.50 millimeters, outperforming the state of the art. Even without additional synthesized data, our method achieved competitive results (Hausdorff Distance: 9.43 millimeters, Chamfer Distance: 2.94 millimeters). Qualitative results demonstrated accurate facial predictions. The analysis revealed slightly higher sensitivity (54.20% compared to 53.30%) and precision (50.20% compared to 49.40%) for engineers compared to medical professionals, though both groups had low specificity, approximately 46%. Statistical tests showed no significant difference in distinguishing Machine Learning-Generated faces from Real Surgical Outcomes, with p-values of 0.567 and 0.256, respectively. Ablation