深度学习在上气道三维分析与正颌术后上气道预测中的应用进展

Advances in Deep Learning for Three-Dimensional Analysis of the Upper Airway and Prediction of Upper Airway Changes after Orthognathic Surgery

摘要: 上气道形态评估对牙颌面畸形诊疗及正颌外科预后至关重要。锥形束计算机断层扫描(Cone Beam Computed Tomography, CBCT)结合三维重建技术可实现精确的气道分析。然而，手动分割方法存在耗时且主观性强的问题。近年来，深度学习(Deep Learning, DL)尤其是卷积神经网络(Convolutional Neural Network, CNN)的引入，显著提升了CBCT上气道自动化分析的效率与准确性。本文综述了DL在气道辅助诊断、自动分割及参数测量中的应用，在辅助诊断分级、高精度气道分割、自动测量气道参数等方面DL均展现出优于传统方法的性能。在正颌术后气道预测方面，DL模型初步实现了术后气道容积与流体力学特征的精准预测，为个性化手术方案设计提供支持。然而，当前研究仍面临数据集规模有限、模型可解释性不足、多模态融合欠缺及计算资源依赖等挑战。未来需构建大规模多中心数据库，发展可解释性轻量级模型，并加强临床验证，以推动智能气道分析工具在临床的转化应用。

Abstract: Assessment of upper airway morphology is crucial for the diagnosis and treatment of dentofacial deformities and for predicting the prognosis of orthognathic surgery. Cone-beam computed tomography (CBCT) combined with 3D reconstruction technology enables precise airway analysis. However, manual segmentation methods are time-consuming and highly subjective. In recent years, the introduction of deep learning (DL), particularly convolutional neural networks (CNNs), has significantly improved the efficiency and accuracy of automated upper airway analysis using CBCT. This review summarizes the applications of DL in airway-assisted diagnosis, automatic segmentation, and parameter measurement. DL has demonstrated superior performance compared to traditional methods in areas such as assisted diagnostic grading, high-precision airway segmentation, and automatic measurement of airway parameters. Regarding post-orthognathic surgery airway prediction, DL models have preliminarily achieved accurate prediction of postoperative airway volume and hemodynamic characteristics, providing support for the design of personalized surgical plans. However, current research still faces challenges such as limited dataset size, insufficient model interpretability, lack of multimodal fusion, and dependence on computational resources. Future efforts should focus on establishing large-scale, multicenter databases, developing lightweight, interpretable models, and strengthening clinical validation to advance the clinical translation of intelligent airway analysis tools.