眼底荧光血管造影图片中无灌注区的人工智能识别  

Artificial intelligence recognition of non-perfusion areas in fundus fluorescence angiography images

作　　者：巩迪 赵丽娟[2] 张利[1] 虞欣 陈宜[1] GONG Di;ZHAO Li-juan;ZHANG Li(Department of Ophthalmology,China-Japan Friendship Hospital,Beijing 100029,China)

机构地区：[1]中日友好医院眼科,北京100029 [2]山东省东营市胜利油田中心医院眼科,山东东营257099 [3]北京石油化工学院人工智能研究院,北京102624

出　　处：《中日友好医院学报》2024年第6期323-326,F0002,共5页Journal of China-Japan Friendship Hospital

基　　金：北京科技新星计划资助(20220484158)。

摘　　要：目的:探讨人工智能识别眼底荧光血管造影图片中无灌注区的能力。方法:收集中日友好医院眼科2018年3月—2023年6月期间的眼底荧光血管造影检查结果,1000例患者中筛查出无灌注患者347例,获取无灌注区的眼底荧光血管造影照片622张。1名具有10年临床经验的眼科医生利用RectLabel标注软件中的多边形工具对其中眼底荧光血管造影图片进行手工标注。本研究的深度学习模型采用Vision Transformer与卷积神经网络相融合的总体网络结构,采用渐进式迁移学习策略和加权损失函数强化模型的识别能力。结果:共标注991个无灌注区作为“正样本”,同时在荧光造影图片上无灌注区以外区域,随机裁切一定数量的、大小相同的矩形区域,作为“负样本”。考虑到视网膜无灌注区实际发病率以及在临床上的比例,最终正负样本的采样比例约为1∶9,负样本的数量为8878个。得到的预测灵敏度为98.7%,特异度为95.9%,约登指数为94.6%,查准率为73.3%,F分数为77.9%,准确度为96.9%。结论:本研究的深度学习模型识别眼底荧光血管造影中的无灌注区的能力较强,敏感性和特异性均较高。Objective:To explore the ability of artificial intelligence to recognize non-perfusion areas in fundus fluorescence angiography(FFA)images.Methods:We collected the results of fundus fluorescence angiography in our Ophthalmology Department from March 2018 to June 2023.Among 1000 patients,347 whose fluorescence angiography images of fundus with non-perfusion areas were screened,and 622 FFA photos were obtained with non-perfusion areas.An ophthalmologist with 10 years of clinical experience manually annotated the images using the polygon tool in RectLabel annotation software.The model adopts an overall network structure that combines Vision Transformer and Convolutional Neural Network.Meanwhile,the model adopts a progressive transfer learning strategy and weighted loss function to enhance the recognition ability of the model.Results:A total of 991 areas with non-perfusion were labeled as "positive samples",and a certain number of rectangular areas of the same size were randomly cut as "negative samples" in the areas outside the nonperfusion areas on the fluorescence imaging.Considering the actual incidence rate and clinical proportion of the retinal non-perfusion area,the final sampling ratio of positive and negative samples was 1∶9,and the number of negative samples was 8878.The obtained prediction sensitivity was 98.7%,specificity was 95.9%,Jordan index was 94.6%,precision was 73.3%,F-score was 77.9%,and accuracy was 96.9%.Conclusion:The model used in this study has a strong ability to identify non-perfused areas in fundus fluorescence angiography,with high sensitivity and specificity.

关 键 词：眼底荧光血管造影 无灌注区 人工智能 

分 类 号：R774.1[医药卫生—眼科]