OTOLens：用于评估鼓膜和中耳病变的移动应用程序的开发与验证

文献信息

中文摘要

背景/目的： 由于解剖结构复杂、表现重叠以及专业设备可及性有限，在初级保健中诊断鼓膜和中耳病变具有挑战性。我们开发并外部验证了OTOLens，这是一款基于智能手机的移动应用程序，使用深度学习对耳镜图像进行分类。
方法： 在这项前瞻性开发与验证研究中，OTOLens将三个基于ResNet的图像分类器（OME、TMP和MYR）与一个在PCA降维图像特征上运行的平行梯度提升表格分类器相结合，通过按病种取最大值和0.5决策阈值进行整合（Apple Core ML/Create ML）；该系统在1011张标注耳镜图像上进行训练，涵盖五个类别：正常鼓膜（n = 251）、伴渗出液的中耳炎（OME；n = 227）、鼓膜炎（MYR；n = 210）、其他病变（n = 197）和鼓膜穿孔（TMP；n = 126），由三位接受过专科培训的耳鼻咽喉科医师通过多数共识进行标注。外部验证使用180张前瞻性采集的图像，由三位耳鼻咽喉科医师独立审阅。性能指标包括敏感性、特异性、阳性预测值、F1分数和AUROC；评分者间信度采用Fleiss' κ评估；应用程序一致性采用Cohen's κ评估。
结果： 内部测试显示各类别准确率为58-81%，AUROC最高达0.89。外部验证达到52.8%的总体准确率（Cohen's κ = 0.417；中等一致性）。各类别敏感性为：TMP 86.7%、OME 48.9%、MYR 40.0%、正常鼓膜35.6%；正常鼓膜和MYR的特异性为100%。专家评分者间信度为显著（Fleiss' κ = 0.681，95% CI：0.619-0.738）。
结论： OTOLens对TMP达到了高敏感性，支持其作为初级保健分诊辅助工具的潜力。这一单一较强结果并不构成广泛的临床效用；正常鼓膜、OME和MYR的低敏感性以及中等的OME阳性预测值表明，该应用程序无法可靠地评估这些类别，不应单独使用。由于域偏移、类别不平衡和TMP代表性有限，泛化性受到限制；在临床部署之前，需要使用扩展数据集进行多中心验证。

英文摘要

BACKGROUND/OBJECTIVE: Diagnosing tympanic membrane and middle ear pathology in primary care is challenging due to anatomical complexity, overlapping presentations, and limited access to specialized equipment. We developed and externally validated OTOLens, a smartphone-based mobile application classifying otoscopic images using deep learning.
METHODS: In this prospective development and validation study, OTOLens combined three ResNet-based image classifiers (OME, TMP, and MYR) with a parallel gradient-boosted tabular classifier operating on PCA-reduced image features, integrated by a per-condition maximum and a 0.5 decision threshold (Apple Core ML/Create ML); the system was trained on 1011 annotated otoscopic images across five categories: normal tympanic membrane (n = 251), otitis media with effusion (OME; n = 227), myringitis (MYR; n = 210), other pathology (n = 197), and tympanic membrane perforation (TMP; n = 126), labeled by three fellowship-trained otolaryngologists via majority consensus. External validation used 180 prospectively acquired images reviewed independently by three otolaryngologists. Performance metrics included sensitivity, specificity, positive predictive value, F1 score, and AUROC; inter-rater reliability was assessed by Fleiss' κ; application agreement by Cohen's κ.
RESULTS: Internal testing demonstrated class-wise accuracy of 58-81% with AUROC up to 0.89. External validation achieved 52.8% overall accuracy (Cohen's κ = 0.417; moderate agreement). Class-specific sensitivity was 86.7% for TMP, 48.9% for OME, 40.0% for MYR, and 35.6% for normal tympanic membrane; specificity was 100% for normal tympanic membrane and MYR. Expert inter-rater reliability was substantial (Fleiss' κ = 0.681, 95% CI: 0.619-0.738).
CONCLUSIONS: OTOLens achieved high sensitivity for TMP, supporting its potential as a primary care triage aid. This single stronger result does not constitute broad clinical utility; the low sensitivities for normal TM, OME, and MYR and the modest OME PPV indicate that the application cannot reliably assess these categories and should not be used in isolation. Generalizability was constrained by domain shift, class imbalance, and limited TMP representation; multicenter validation with expanded datasets is required prior to clinical deployment.