人工智能在慢性鼻窦炎中的应用：从现状到未来

文献信息

中文摘要

慢性鼻窦炎（CRS）是一种常见的、生物学上异质性的疾病，其诊断、内型分型、治疗选择以及手术决策均仍不完善。在短短十多年间，应用于CRS的人工智能（AI）已从经典回归发展到机器学习、深度学习，以及最近的计算机视觉和生成模型，产生的工具现已触及整个诊疗路径。本综述综合了在四项当前能力和两个新兴前沿方面的进展。在诊断方面，机器学习可从就诊前数据预测可能的CRS，而卷积网络可解读CT、鼻内镜和组织病理学，并无创地推断嗜酸性粒细胞内型。在内型分型方面，监督式、非监督式和潜因子模型，连同基于图像的深度学习，可解析CRS中分子定义的亚群，反复提示IL-5以及混合型2型/中性粒细胞性炎症。在治疗方面，机器学习可阐明药物治疗反应，并识别生物制剂获益的分子预测因子。在手术方面，模型可预测内镜鼻窦手术的成功和疾病复发，并集中于哮喘、影像学鼻窦混浊以及嗜酸性粒细胞或中性粒细胞性炎症。展望未来，多模态基础模型、联邦学习和计算机视觉引导的术中导航有望实现整合的、患者特异性的决策支持。我们还权衡了AI在CRS诊疗中的风险和局限性，并概述了未满足的需求，包括就AI如何开发和评估达成国际共识的理由。在各个领域，复杂算法并不总是优于简约模型，而外部验证仍然是临床转化的主要障碍。

英文摘要

Chronic rhinosinusitis (CRS) is a common, biologically heterogeneous disease in which diagnosis, endotyping, therapy selection, and surgical decision-making all remain imperfect. Over little more than a decade, artificial intelligence (AI) applied to CRS has advanced from classical regression to machine learning, deep learning, and, most recently, computer vision and generative models, producing tools that now touch the entire care pathway. This review synthesizes that progress across four present-day capabilities and two emerging frontiers. In diagnosis, machine learning predicts probable CRS from pre-visit data while convolutional networks interpret CT, nasal endoscopy, and histopathology and infer eosinophilic endotype noninvasively. In endotyping, supervised, unsupervised, and latent-factor models, together with image-based deep learning, resolve molecularly defined subsets in CRS, repeatedly implicating IL-5 and mixed type 2/neutrophilic inflammation. In therapeutics, machine learning clarifies medical-therapy response and identifies molecular predictors of biologic benefit. In surgery, models forecast endoscopic sinus surgery success and disease recurrence, converging on asthma, radiographic sinus opacification, and eosinophilic or neutrophilic inflammation. Looking ahead, multimodal foundation models, federated learning, and computer-vision-guided intraoperative navigation promise integrated, patient-specific decision support. We also weigh the risks and limitations of AI in CRS care and outline unmet needs, including the case for international consensus on how AI is developed and evaluated. Across domains, complex algorithms do not consistently outperform parsimonious models, and external validation remains the principal barrier to clinical translation.