深度学习在头影测量中的应用研究进展

Research progress on the application of deep learning in cephalometric analysis

头影测量是正畸和正颌外科诊疗过程中普遍采用的关键技术。常规的手工头影测量和半自动头影测量依赖于医师手工定点,存在耗时过长、精确度和稳定性欠佳等问题。深度学习(deep learning,DL)是人工智能(artificial intelligence,AI)领域最热门的技术方法之一,具有强大的图像识别能力。在头影测量领域,目前已有较多研究致力于采用DL实现头影测量的自动化,包括自动定点、自动诊断分析、颈椎成熟度分析、腺样体肥大自动诊断和气道分析。研究显示,DL能有效提高工作效率,其准确率多已达到80%以上,与金标准之间的差异多在临床可接受范围内,具有良好的应用前景。然而,当前此类研究仅限于自动定点的实现,其数据集往往存在数据量不够丰富、数据范围局限等问题。后期研究应进一步扩展研究范围、提升算法的同时注意数据集的丰富程度,并将DL技术与其他AI技术适当结合,从而进一步提高算法的准确性、稳定性和普适性。

In orthodontic and orthognathic practice,cephalometric analysis is an integral tool throughout the clinical process.However,as landmark identification is still unautomated,both the conventional and semiautomated approaches are open to considerable subjectivity and could be timeconsuming for inexperienced clinicians.Deep learning(DL),a stateoftheart artificial intelligence(AI)technique,is highly effective in image recognition.In recent years,many studies have focused on the application of DL in cephalometric analysis,including automated landmark detection,automated diagnosis,cervical vertebral maturation stage determination,adenoid hypertrophy analysis and upper airway identification.Studies show that DL can effectively improve the efficiency of cephalometric analysis.In most studies,the accuracy of DL can reach more than 80%,and its difference from the gold standard is clinically acceptable,demonstrating good potential for future applications.However,most studies are limited to landmark detection,and the broadness and richness of the training dataset are limited.Future studies should broaden the research scope,improve the algorithm,elevate the richness of the datasets,and combine DL with other AI algorithms to improve its accuracy,stability and generalizability.