人工智能在儿童腺样体肥大筛查与诊断中的 进展

Advances in Artificial Intelligence for the Screening and Diagnosis of Pediatric Adenoid Hypertrophy

摘要: 腺样体肥大在儿童时期发病率高达34%，易造成打鼾、反复感染及阻塞性睡眠呼吸暂停等不良后果，影响儿童生长发育。传统的诊断方法(临床评分、鼻内镜、鼻咽侧位片、CBCT、MRI等)已形成体系，但受侵袭性、辐射暴露、技术敏感性与可及性限制。近年来，人工智能在多模态诊断中具有潜力：在侧位片上实现了A/N比自动测量与分级，阅片效率得到显著提升；在鼻内镜分级中借助弱标注与对比学习提高诊断稳健性；CBCT三维分割中可获得较高Dice值，并可输出体积与最窄横截面积等定量指标；MRI的关键点定位与测量初步可行，但三维信息尚未得到充分利用；面部图像与心肺音诊断作为无创、无辐射的筛查手段具有前景，但样本量与外部验证仍不足。面向临床落地，下一步需强化多中心外部验证、标准化采集与标注，重视可解释性与隐私合规(如联邦学习)、增强模型校准与决策曲线及成本–效果评估，并探索多模态融合从“辅助诊断”迈向“治疗决策支持”(如手术指征与路径选择)，以提升临床诊断一致性与决策价值。

Abstract: Adenoid hypertrophy has a pediatric prevalence of up to 34% and is closely associated with snoring, recurrent infections, and obstructive sleep apnea, adversely affecting growth and development. Conventional diagnostic methods (clinical scoring, nasal endoscopy, nasopharyngeal lateral radiograph, CBCT, MRI) are well established but constrained by invasiveness, ionizing radiation, operator variability, and limited accessibility. Artificial intelligence shows promise across modalities. On lateral radiographs, automated A/N ratio measurement and grading improve efficiency and consistency. In nasal endoscopy, weak supervision and contrastive learning enhance robustness. CBCT enables 3D segmentation with high Dice scores and quantitative metrics such as volume and minimal cross-sectional area. MRI supports keypoint localization and measurement, but current work underuses full 3D information. Facial images and cardiopulmonary sounds offer noninvasive, radiation-free screening options, though evidence is limited and external validation is scarce. For clinical adoption, priorities include multicenter external validation, domain adaptation, and standardized acquisition and annotation. Explainability and privacy compliance—such as federated learning—are essential. Models should be well-calibrated and assessed with decision curves and cost-effectiveness. Multimodal fusion can help move from assisted diagnosis to treatment decision support, including surgical indications and pathway selection, and improve real-world decision quality.