基于多光谱眼底成像开发的人工智能对糖尿病视网膜病早期体征的识别  

Research on Recognition of Early Signs of Diabetic Retinopathy by Artificial Intelligence Based on Multispectral Fundus Imaging

作　　者：马菲妍 张彩霞 冬雪川 MA Fei-yan;ZHANG Cai-xia;DONG Xue-chuan(The Second Hospital of Hebei Medical University,Shijiazhuang 050005,China;Shenzhen New Industry Ophthalmology New Technology Co.,Ltd.,Shenzhen 518055,China)

机构地区：[1]河北医科大学第二医院,河北石家庄050005 [2]深圳市新产业眼科新技术有限公司,广东深圳518055

出　　处：《吉林医学》2022年第2期386-390,共5页Jilin Medical Journal

基　　金：河北省2020年度医学科学研究课题计划[项目编号:20200069];深圳市科技研发资金-深科技创新[2019]33号[项目编号:JSGG20180507182010237]。

摘　　要：目的:探索基于多光谱眼底成像开发的人工智能在糖尿病视网膜病早期诊断上的效能。方法:采用诊断试验的研究方法,以200张经过专家标定的多光谱眼底图像作为阅片标注的参考标准,比较AI组、高年资眼科医师组、低年资眼科医师组及内分泌医师组的诊断一致性和阅片速度。不同阅片者的阅片结果和参考标准的比较采用加权kappa系数进行评价,AI组和医师组的比较以Kendall系数进行评价。AI系统和各医师组的单张平均阅片时间比较采用重复测量方差Bonferroni法分析。结果:对于各组的阅片一致性,AI组和高年资眼科医师组、低年资眼科医师组、内分泌医师组相比,Kendall协调系数为0.957、0.947和0.926,差异有统计学意义(P<0.01),诊断结果基本一致,内分泌医师组和低年资眼科医师组协调系数依次低于高年资眼科医师组,但差异无显著统计学意义(P>0.05)。各组平均阅片时间的差异具有统计学意义(F=3 220.879,P<0.01),AI组的阅片时间明显少于各医师组。结论:通过人工智能和多光谱眼底成像技术的结合能够提升对糖尿病视网膜病的筛查效能,降低开发难度,有利于不同资质的医师,尤其是全科医师和年轻医师快速掌握本病的诊断和筛查。Objective To explore the effectiveness of artificial intelligence(AI)deep learning technology in the screening and detecting Diabetic Retinopathy(DR).Method Using the research method of diagnostic tests,200 multispectral fundus images calibrated by experts were used as the reference standard for reading and labeling,compare the AI group,senior ophthalmologist group,junior ophthalmologist group and the endocrinologist group Diagnostic consistency and speed of reading.The comparison between the reading results of different readers and the reference standard was evaluated by the weighted kappa coefficient,the comparison between the AI group and the physician group was evaluated by the Kendall coefficient.The comparison of the average reading time of a single image between the AI system and each physician group was analyzed by the repeated measures variance variance Bonferroni method.Results The Kendall coordination coefficients between the AI group and the senior ophthalmologist group,junior ophthalmologist group,endocrinologist group were 0.957,0.947 and 0.926,the difference was statistically significant(P<0.01),the diagnosis level was close;the endocrinologist group and the junior ophthalmologist group were lower than those of the AI group and the senior ophthalmologist group,but there was no significant statistical difference(P>0.05).The difference in the average reading time of each group was statistically significant(F=3220.879,P<0.01).Conclusion The combination of artificial intelligence and multi-spectral fundus imaging technology can improve the screening efficiency of diabetic retinopathy,reduce the difficulty of development,and help doctors of different qualifications,especially general practitioners and young doctors,to quickly grasp the diagnosis and diagnosis of this disease.

关 键 词：多光谱眼底成像 人工智能 糖尿病视网膜病 

分 类 号：R587.2[医药卫生—内分泌] R774.1[医药卫生—内科学]