基于深度学习的早产儿视网膜病变的临床辅助诊断

中华实验眼科杂志2019年8月第37卷第8期Chin J Exp Ophthalmol,August2019,Vol.37,No.8•647・

•临床研究•

基于深度学习的早产儿视网膜病变的

临床辅助诊断

童妍卢苇徐阳涛李莹王晓玲陈长征沈吟

武汉大学人民医院眼科中心430060

通信作者：沈吟,Email:yinshen@

【摘要】目的提出一种基于深度学习的早产儿视网膜病变(ROP)智能辅助诊断系统,并评估其在临

床上的应用价值。方法采集武汉大学人民医院眼科中心2009年7月至2016年12月行早产儿眼底筛査

的38895张图像构建眼底图像大规模数据集，由10名眼科医生进行标注，建立深度学习网络，通过对模型的

训练实现ROP的自动诊断,评估该算法自动筛査ROP分期、分区及附加病变的性能和准确率。结果深度

学习智能诊断系统对ROP分期及其附加病变、视盘、黄斑及激光治疗瘢痕检测的平均准确率为0.931；其中检

测分界线(I期)准确率为0.876,视网膜Il#(n期)为0.942,膜蜡伴血管扩张(HI期)为0.968；视网膜不完全

脱离(IV期)为0.998,视网膜完全脱离(V期)为0999；血管迂曲扩张(附加病变)为0.896,视盘为0.954,黄

斑为0.781,激光治疗瘢痕为0.974。结论基于深度学习算法的ROP的疾病分期和附加病变的诊断准确

率高,可用于ROP的临床辅助诊断和筛査。

【关键词】人工智能；深度学习；早产儿视网膜病变；辅助诊断

基金项目：国家重点研发计划“政府间国际科技创新合作/港澳台科技创新合作”重点专项项目

(2017YFE0103400)

D01:10.3760/cma.j.issn.2095-0160.2019.08.011

Automated assisted clinical diagnosis of retinopathy of prematurity based on deep learning

Tong Yan,Lu Wei,Xu Yangtao,Li Ying,Wang Xiaoling,Chen Changzheng,Shen Yin

Eye Center,Renmin Hospital of Wuhan University,Wuhan430060,China

Corresponding author:Shen Yin,Email:yinshen@

[Abstract]Objective To evaluate the application value of an intelligent fundus assisted diagnosis system for

detecting retinopathy of prematurity(ROP)based on deep learning.Methods A total of38895fundus images for

premature infants screening were collected from Renmin Hospital of Wuhan University Eye Center and were labeled by

10licensed ophthalmologists.A deep learning network model was established to acquire automatic classification of

disease stages and plus disease.The accuracy,sensitivity and specificity of the algorithm were calculated to evaluate

the performance of the artificial intelligence system for ROP automatic diagnosis.This study protocol was approved by

Ethic Committee of Renmin Hospital of Wuhan University(No.WDRY2019-K032).Written informed consent was

obtained from the guardians of the children before entering the study cohort.Results The intelligent system

achieved an accuracy of0.931.Specifically,the accuracies in detecting demarcation line(stage I)was0.876,ridge

(stage II)was0.942,ridge with extra retinal fibrovascular(stage皿)was0.968,subtotal retinal detachment(stage

IV)was0.998,total retinal detachment(stage V)was0.999,vascular tortuosity and dilatation(plus disease)was

0.896,optic disc was0.954,macular was0.781,and laser scars were0.974,respectively.Conclusions Deep

learning algorithm can detect the stages and plus disease of ROP with excellent accuracy,and it provides the feasibility

of applying the algorithm for ROP automated screening in clinical.

[Key words]Artificial intelligence；Deep learning；Retinopathy of prematurity；Assisted diagnosis

Fund Program:International Science&Technology Cooperation Program of China(2017YFE0103400)

DOI:10.3760/cma.j.issn.2095-0160.2019.08.011

早产儿视网膜病变（retinopathy of prematurity,早产儿及低出生体质量新生儿的存活率不断提高，R0P）是世界首位儿童不可逆性致盲眼病口勺。随着R0P的发生率也随之升高。R0P发病隐匿，呈进行性