ABSTRACT

Purpose: To propose a deep-learning-based approach to automatically and objectively evaluate morphologic eyelid features using two-dimensional(2D) digital photographs and to assess the agreement between automatic and manual measurements.

Materials and Methods: The 2D photographs of 1378 normal Asian participants (2756 eyes) were included for training, validating and testing the cornea and eyelid segmentation network. Margin reflex distance 1 (MRD1) and margin reflex distance 2 (MRD2) of 406 eyes from 203 participants were manually evaluated by 3 ophthalmologists and the photographs of 406 eyes were measured automatically for 8 morphologic eyelid features. The Spearman’s correlation coefficient, intra-class correlation coefficient (ICC) and Bland-Altman analyses were used to determine the agreement between manual and automatic MRDs.

Results: The dice coefficient was 0.922 and 0.974 for eyelid and cornea segmentation, respectively. A strong correlation was shown between manually and automatically measured MRD1 (r = 0.993, ICC = 0.996) and MRD2 (r = 0.950, ICC = 0.974). Bland-Altman analyses also showed excellent reliability with bias being 0.04 mmbetween automated and manual MRD1 measurements and 0.06 mm for MRD2. Automatically measured 8 features (MRD1, MRD2, palpebral fissure, medial area, lateral area, cornea area, upper and lower eyelid lengths) were found to be increased with age and peaked around the age range of 21 to 30 years.

Conclusions: The proposed novel integrative analysis scheme was comparable with human performance. The approach with excellent reliability and reproductivity showed great potential for automated diagnosis and remote monitoring of eyelid-related diseases.

中文翻译：

摘要

目的：提出一种基于深度学习的方法，使用二维 (2D) 数码照片自动和客观地评估形态学眼睑特征，并评估自动和手动测量之间的一致性。

材料和方法：包括 1378 名正常亚洲参与者（2756 只眼睛）的 2D 照片，用于训练、验证和测试角膜和眼睑分割网络。由 3 名眼科医生手动评估来自 203 名参与者的 406 只眼睛的边缘反射距离 1 (MRD1) 和边缘反射距离 2 (MRD2)，并自动测量 406 只眼睛的 8 个形态眼睑特征的照片。Spearman 相关系数、类内相关系数 (ICC) 和 Bland-Altman 分析用于确定手动和自动 MRD 之间的一致性。

结果：眼睑和角膜分割的骰子系数分别为 0.922 和 0.974。手动和自动测量的 MRD1 (r = 0.993, ICC = 0.996) 和 MRD2 (r = 0.950, ICC = 0.974) 之间显示出很强的相关性。Bland-Altman 分析还显示了出色的可靠性，自动和手动 MRD1 测量的偏差为 0.04 mm，MRD2 的偏差为 0.06 mm。发现自动测量的 8 个特征（MRD1、MRD2、睑裂、内侧区域、外侧区域、角膜区域、上下眼睑长度）随着年龄的增长而增加，并在 21 至 30 岁的年龄范围内达到峰值。

结论：所提出的新型综合分析方案可与人类表现相媲美。该方法具有出色的可靠性和再现性，在眼睑相关疾病的自动诊断和远程监测方面显示出巨大的潜力。