Cai Jie, Xiu Tianyu, Song Yuliang, Fan Xuwei, Wu Jianghao, Tuohuti Aikebaier, Hu Yifan, Chen Xiong
Department of Otorhinolaryngology, Head and Neck Surgery, Zhongnan Hospital of Wuhan University, Wuhan, 430000, People's Republic of China.
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, 430079, People's Republic of China.
Nat Sci Sleep. 2024 Dec 27;16:2243-2256. doi: 10.2147/NSS.S492146. eCollection 2024.
This study aims to develop a deep learning methodology for quantitative assessing adenoid hypertrophy in nasopharyngoscopy images and to investigate its correlation with the apnea-hypopnea index (AHI) in pediatric patients with obstructive sleep apnea (OSA).
A total of 1642 nasopharyngoscopy images were collected from pediatric patients aged 3 to 12 years. After excluding images with obscured secretions, incomplete adenoid exposure, 1500 images were retained for analysis. The adenoid-to-nasopharyngeal (A/N) ratio was manually annotated by two experienced otolaryngologists using MATLAB's imfreehand tool. Inter-annotator agreement was assessed using the Mann-Whitney -test. Deep learning segmentation models were developed with the MMSegmentation framework, incorporating transfer learning and ensemble learning techniques. Model performance was evaluated using precision, recall, mean intersection over union (MIoU), overall accuracy, Cohen's Kappa, confusion matrices, and receiver operating characteristic (ROC) curves. The correlation between the A/N ratio and AHI, derived from polysomnography, was analyzed to evaluate clinical relevance.
Manual evaluation of adenoid hypertrophy by otolaryngologists (p=0.8507) and MATLAB calibration (p=0.679) demonstrated high consistency, with no significant differences. Among the deep learning models, the ensemble learning-based SUMNet outperformed others, achieving the highest precision (0.9616), MIoU (0.8046), overall accuracy (0.9182), and Kappa (0.87). SUMNet also exhibited superior consistency in classifying adenoid sizes. ROC analysis revealed that SUMNet (AUC=0.85) outperformed expert evaluations (AUC=0.74). A strong positive correlation was observed between the A/N ratio and AHI, with the correlation coefficients for SUMNet-derived ratios ranging from r=0.9052 (tonsils size+1) to r=0.4452 (tonsils size+3) and for expert-derived ratios ranging from r=0.4590 (tonsils size+1) to r=0.2681 (tonsils size+3).
This study introduces a precise and reliable deep learning-based method for quantifying adenoid hypertrophy and addresses the challenge posed limited sample sizes in deep learning applications. The significant correlation between adenoid hypertrophy and AHI underscores the clinical utility of this method in pediatric OSA diagnosis.
本研究旨在开发一种深度学习方法，用于定量评估鼻咽喉镜图像中的腺样体肥大，并研究其与阻塞性睡眠呼吸暂停（OSA）儿科患者的呼吸暂停低通气指数（AHI）之间的相关性。
共收集了1642例3至12岁儿科患者的鼻咽喉镜图像。在排除有分泌物遮挡、腺样体暴露不完整的图像后，保留1500张图像用于分析。两名经验丰富的耳鼻喉科医生使用MATLAB的imfreehand工具手动标注腺样体与鼻咽（A/N）比值。使用曼-惠特尼检验评估标注者间的一致性。使用MMSegmentation框架开发深度学习分割模型，纳入迁移学习和集成学习技术。使用精度、召回率、平均交并比（MIoU）、总体准确率、科恩kappa系数、混淆矩阵和受试者工作特征（ROC）曲线评估模型性能。分析来自多导睡眠图的A/N比值与AHI之间的相关性，以评估临床相关性。
耳鼻喉科医生对腺样体肥大的手动评估（p = 0.8507）和MATLAB校准（p = 0.679）显示出高度一致性，无显著差异。在深度学习模型中，基于集成学习的SUMNet表现优于其他模型，达到了最高精度（0.9616）、MIoU（0.8046）、总体准确率（0.9182）和kappa系数（0.87）。SUMNet在分类腺样体大小方面也表现出更高的一致性。ROC分析显示SUMNet（AUC = 0.85）优于专家评估（AUC = 0.74）。观察到A/N比值与AHI之间存在强正相关，SUMNet得出的比值的相关系数范围为r = 0.9052（扁桃体大小+1）至r = 0.4452（扁桃体大小+3），专家得出的比值的相关系数范围为r = 0.4590（扁桃体大小+1）至r = 0.2681（扁桃体大小+3）。
本研究引入了一种基于深度学习的精确可靠方法来量化腺样体肥大，并解决了深度学习应用中样本量有限带来的挑战。腺样体肥大与AHI之间的显著相关性突出了该方法在儿科OSA诊断中的临床实用性。
