军事知识服务平台
人工智能用于Fuchs角膜内皮营养不良严重程度分级的可行性研究
Assessing the feasibility of artificial intelligence for severity grading in fuchs endothelial corneal dystro-phy
摘要 目的 研究人工智能应用于Fuchs角膜内皮营养不良(Fuchs endothelial corneal dystrophy,FECD)严重程度分级的可行性与临床价值.设计 横断面研究.研究对象2022年1月至2025年6月就诊于北京大学第三医院眼科门诊的FECD患者245例.方法 采集和筛选605张FECD患者的共聚焦显微镜图像,人工标记Guttae的边界形成金标准,建立并测试U-Net++神经网络,联合分水岭算法作为后处理.建立以U-Net为代表的9种神经网络作为对照组.主要指标采用准确率、灵敏度、特异度、交并比(IoU)、Dice系数和修正Hausdorff距离(MHD)评价分割模型的准确性,评价模型输出的Guttae参数(面积占比、密度和平均大小)的准确性.结果 245例患者的605张共聚焦显微镜图像中,U-Net++模型的Dice系数为 0.7676,IoU为 0.6229,准确率为 0.9670,敏感度为 0.7923,特异度为 0.9799,MHD 为 33.99 px,AUC为0.9758;Guttae面积占比、密度和平均大小的平均相对误差分别为0.453、0.280、0.222.以上表现整体优于多种对照模型.结论 共聚焦显微镜图像的深度学习模型能有效检测FECD的病灶区域并输出多种Guttae参数,用于FECD的自动化严重程度分级,有助于推进FECD的精准诊疗.
Abstract:
Objective To investigate the feasibility and clinical value of applying artificial intelligence for severity grading of Fuchs endothelial corneal dystrophy(FECD).Design Cross-sectional study.Participants 245 patients with FECD who visited the Ophthalmology Clinic of Peking University Third Hospital between January 2022 and June 2025.Method A total of 605 confocal microscopy images from FECD patients were col-lected and screened.The boundaries of guttae were manually annotated to establish the gold standard.A U-Net++neural network combined with a watershed algorithm as post-processing was developed and tested.Nine neural networks,including U-Net,were constructed as control models.Main Outcome Measures The segmentation accuracy was evaluated using accuracy,sensitivity,specificity,Intersection over Union(IoU),Dice coefficient,and modified Hausdorff distance(MHD).The accuracy of the guttae parameters(area ratio,density,and average size)output by the model was also assessed.Results Among 605 confocal microscope images of 245 patients,the U-Net++model achieved a Dice coefficient of 0.7676,IoU of 0.6229,accuracy of 0.9670,sensitivity of 0.7923,specificity of 0.9799,MHD of 33.99 px,and AUC of 0.9758.The mean relative errors for guttae area ratio,density,and average size were 0.453,0.280,and 0.222,respectively.Overall,these results outperformed multiple control models.Conclusions The deep learning model on confocal microscope images can effectively detect lesion regions in FECD and output multiple guttae parameters for automated severity grading,contributing to the advancement of precise diagnosis and management of FECD.
刊 名 眼科 ISTIC
年,卷(期): 2026, 35(4)
机标分类号: TP391.41 R329.2 Q246
在线出版日期: 2026年8月20日