基于锥形束计算机断层扫描的深度学习增强模型用于筛查睡眠呼吸障碍风险
Cone-beam computed tomography-based deep learning-augmented model for screening sleep-disordered breathing risk.
文献信息
| PMID | 42836807 |
|---|---|
| 原文 | 在 PubMed 查看原文 ↗ |
| 发表日期 | 2026 |
| 作者 | Min Yu |
| 作者单位 | Department of Orthodontics, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China; Center for Oral Therapy of Sleep Apnea, Peking University Hospital of Stomatology, Beijing, People's Republic of China; National Center for Stomatology, Beijing, People's Republic of China; Sanya Stomatology Center, Peking University School of Stomatology Sanya Division, Sanya City, People's Republic of China. |
| 期刊 | American journal of orthodontics and dentofacial orthopedics : official publication of the American Association of Orthodontists, its constituent societies, and the American Board of Orthodontics |
| SCI 分区 | Q1 |
| IF | 4.1 |
| 研究类型 | AI/ML · 临床 |
| 所属专科 | 鼻科 |
中文摘要
引言： 睡眠呼吸障碍（SDB）是一种常见但诊断不足的疾病，其诊断程序复杂，构成了巨大的全球健康负担。颅面解剖结构，如下颌后缩和上气道狭窄，在亚洲人群的SDB中起重要作用，凸显了正畸医生在SDB风险筛查中的潜在作用。本研究旨在开发一种基于锥形束计算机断层扫描的深度学习增强模型，用于成人SDB风险筛查。
方法： 训练队列包括110例SDB患者和110例性别匹配的非打鼾对照，而独立测试队列由来自另一中心的20例SDB患者和20例对照组成。将预训练的卷积神经网络应用于头颅侧位片，并使用锥形束计算机断层扫描衍生的上气道参数训练多种机器学习模型。使用随机森林分类器构建了一个将深度学习输出与选定的气道特征相结合的晚期融合框架。应用可解释性和可视化技术以提高可解释性。
结果： 在使用头颅侧位片评估的6个预训练卷积神经网络模型中，ResNet-50达到了最高的曲线下面积（AUC）0.640。其次，当应用于上气道体积参数时，随机森林分类器优于其他机器学习算法，AUC为0.908。晚期融合模型进一步提高了筛查性能，AUC达到0.925，并优于所有单模态模型。可视化分析突出了易患SDB的气道相关特征，特别是在腭咽区域。
结论： 这种深度特征增强融合模型在SDB风险筛查中表现出改进的性能和可解释性。常规获取的正畸影像可能作为一种实用且对正畸医生友好的方法，支持早期识别有SDB风险的患者。
英文摘要
INTRODUCTION: Sleep-disordered breathing (SDB) is a common but underdiagnosed disorder due to complex diagnostic procedures, posing a substantial global health burden. Craniofacial anatomy, such as mandibular retrusion and upper airway narrowing, plays a prominent role in SDB among Asian populations, underscoring the potential role of orthodontists in SDB risk screening. This study aimed to develop a deep learning-augmented, cone-beam computed tomography-based model for SDB risk screening in adults.
METHODS: A training cohort included 110 patients with SDB and 110 sex-matched nonsnoring controls, whereas an independent test cohort comprised 20 patients with SDB and 20 controls from another center. Pretrained convolutional neural networks were applied to lateral cephalograms, and multiple machine learning models were trained using cone-beam computed tomography-derived upper airway parameters. A late-fusion framework that integrated deep learning outputs with selected airway features was constructed using a random forest classifier. Explainability and visualization techniques were applied to improve interpretability.
RESULTS: Among the 6 pretrained convolutional neural network models evaluated using lateral cephalograms, ResNet-50 achieved the highest area under the curve (AUC) of 0.640. Second, when applied to upper airway volume parameters, the random forest classifier outperformed other machine learning algorithms, yielding an AUC of 0.908. The late-fusion model further improved screening performance, achieving an AUC of 0.925 and outperforming all single-modality models. Visualization analyses highlighted airway-related features that predispose to SDB, particularly within the palatopharyngeal region.
CONCLUSIONS: This deep feature-augmented fusion model demonstrates improved performance and interpretability for SDB risk screening. Routinely acquired orthodontic imaging may serve as a practical and orthodontist-friendly approach to support the early identification of patients at risk for SDB.