基于深度学习模型进行屈光不正与玻璃膜疣的相关性分析  

Correlation Analysis between Refractive Error and the Risk of Drusen Based on a Deep Learning Model

作　　者：顾虹 沈肇萌[1] 李志国[1] Hong Gu;Zhaomeng Shen;Zhiguo Li(Ningbo Medical Center,Lihuili Hospital,Ningbo 315040,China)

机构地区：[1]宁波市医疗中心李惠利医院,315040

出　　处：《中华眼视光学与视觉科学杂志》2020年第5期360-366,共7页Chinese Journal Of Optometry Ophthalmology And Visual Science

摘　　要：目的:采用深度学习模型对彩色眼底图像中的玻璃膜疣进行自动判读,并进一步探讨屈光不正与玻璃膜疣的相关关系。方法:病例对照研究。选取2017年1─12月在宁波市医疗中心李惠利医院行健康体检且年龄大于50岁的参与者,共1 035例(2 055眼)的彩色眼底图像,使用DeepSeeNet深度学习模型对彩色眼底图像进行自动分析判读,按玻璃膜疣大小进行分组。同时随机选择392张(19.1%)彩色眼底图像采用人工阅片方法对玻璃膜疣大小进行判读,采用Cohen's Kappa检验比较2种判读方法的一致性。根据电脑验光数据计算等效球镜度(SE),并分为中重度远视(>+3.0 D)、轻度远视(+0.51^+3.0 D)、正视(-0.5^+0.5 D)、轻度近视(-3.0^-0.51 D)、中重度近视(<-3.0 D)。应用Logistic回归模型进行屈光不正与玻璃膜疣的相关性分析。结果:深度学习模型对玻璃膜疣大小判读与人工阅片结果具有高度一致性( κ=0.67, P<0.001)。在矫正性别、年龄等因素后,SE向正方向增加与大玻璃膜疣发生风险相关( OR=1.03,95% CI:1.01~1.04, P<0.001)。中高度近视为玻璃膜疣的保护因素,与正视者相比,中重度近视者发生大玻璃膜疣的风险为0.89(95% CI:0.82~0.97);相反,中高度远视为危险因素,与正视者相比其发生大玻璃膜疣的风险为1.20(95% CI:1.03~1.39)。 结论:屈光不正与玻璃膜疣的发生相关。深度学习作为一种新型技术,除了可以增强医师临床诊断的速度和精准性,同样可以为年龄相关性黄斑变性相关的科研提供线索。Objective:To classify drusen size in color fundus photographs by using an automated deep learning model and to investigate the association between refractive error and drusen.Methods:This was a casecontrol study.There were 2055 color fundus photographs and refractive status were obtained from 1035 participants aged 50 years and older who underwent physical examinations in Ningbo Medical Center,Lihuili Hospital from January 2017 to December 2017.DeepSeeNet,a deep learning mode,was used to detect drusen size from color fundus photographs.Three hundred ninety-two fundus photographs were randomly selected and assessed manually by one retinal specialist.Cohen's kappa was used to evaluate the consistency between the DeepSeeNet and the retinal specialist.The spherical equivalent refraction was calculated as adiopter(D)according to the computer optometry data,and classified as emmetropia(-0.5-+0.5 D),mild myopia(-3.0--0.51 D)or hyperopia(+0.51-+3.0 D),and moderate to severe myopia(<-3.0 D)or hyperopia(>+3.0 D).Statistical analysis was performed using R statistical software,and a logistic regression model was used to analyze the association between refractive error and drusen size.Results:The classification results of the deep learning model were highly consistent with those of the retinal specialist(κ=0.67,P<0.001).After adjustment for confounding factors,the increase in spherical equivalent was associated with an increased risk of large drusen(OR=1.03,95%CI:1.01-1.04,P<0.001).Compared with emmetropia,moderate to high myopia was associated with a lower OR of large drusen(OR=0.89,95%CI:0.82-0.97);moderate to high hyperopia was associated with a higher OR of large drusen(OR=1.20,95%CI:1.03-1.39).Conclusions:Refractive error is associated with the development of drusen.These results indicate that the use of a deep learning model not only enhances the speed and accuracy of clinical diagnosis,but also provides clues for age-related macular disease related scientific research.

关 键 词：屈光不正 黄斑变性 年龄相关 视网膜玻璃膜疣 深度学习 人工智能 

分 类 号：R778.1[医药卫生—眼科] TP311.13[医药卫生—临床医学]