Nusair Omar, Asadigandomani Hassan, Farrokhpour Hossein, Moosaie Fatemeh, Bibak-Bejandi Zahra, Razavi Alireza, Daneshvar Kimia, Soleimani Mohammad
Kittner Eye Center, Department of Ophthalmology, University of North Carolina, Chapel Hill, NC 27517, USA.
Department of Ophthalmology, University of California, San Francisco, CA 94143, USA.
Vision (Basel). 2025 Aug 18;9(3):71. doi: 10.3390/vision9030071.
We evaluated the clinical applications of artificial intelligence models in diagnosing corneal diseases, highlighting their performance metrics and clinical potential. A systematic search was conducted for several disease categories: keratoconus (KC), Fuch's endothelial corneal dystrophy (FECD), infectious keratitis (IK), corneal neuropathy, dry eye disease (DED), and conjunctival diseases. Metrics such as sensitivity, specificity, accuracy, and area under the curve (AUC) were extracted. Across the diseases, convolutional neural networks and other deep learning models frequently achieved or exceeded established diagnostic benchmarks (AUC > 0.90; sensitivity/specificity > 0.85-0.90), with a particularly strong performance for KC and FECD when trained on consistent imaging modalities such as anterior segment optical coherence tomography (AS-OCT). Models for IK and conjunctival diseases showed promise but faced challenges in heterogeneous image quality and limited objective training criteria. DED and tear film models benefited from multimodal data yet lacked direct comparisons with expert clinicians. Despite high diagnostic precision, challenges from heterogeneous data, a lack of standardization in disease definitions, imaging acquisition, and model training remain. The broad implementation of artificial intelligence must address these limitations to improve eye care equity.
我们评估了人工智能模型在角膜疾病诊断中的临床应用，重点介绍了它们的性能指标和临床潜力。针对几种疾病类别进行了系统检索：圆锥角膜（KC）、富克斯内皮性角膜营养不良（FECD）、感染性角膜炎（IK）、角膜神经病变、干眼疾病（DED）和结膜疾病。提取了敏感性、特异性、准确性和曲线下面积（AUC）等指标。在所有疾病中，卷积神经网络和其他深度学习模型经常达到或超过既定的诊断基准（AUC > 0.90；敏感性/特异性 > 0.85 - 0.90），在使用前段光学相干断层扫描（AS - OCT）等一致成像模式进行训练时，KC和FECD的表现尤为突出。IK和结膜疾病的模型显示出前景，但在图像质量异质性和客观训练标准有限方面面临挑战。DED和泪膜模型受益于多模态数据，但缺乏与专家临床医生的直接比较。尽管诊断精度很高，但来自异质数据、疾病定义缺乏标准化、成像采集和模型训练等方面的挑战仍然存在。人工智能的广泛应用必须解决这些限制，以改善眼保健公平性。
