Graduate Institute of Biomedical Informatics, College of Medical Science and Technology, Taipei Medical University, Taipei, Taiwan; International Center for Health Information Technology (ICHIT), Taipei Medical University, Taipei, Taiwan; Research Center of Big Data and Meta-Analysis, Wan Fang Hospital, Taipei Medical University, Taipei, Taiwan.
School of Health Care Administration, Taipei Medical University, Taipei, Taiwan.
Comput Methods Programs Biomed. 2020 Jul;191:105320. doi: 10.1016/j.cmpb.2020.105320. Epub 2020 Jan 16.
Diabetic retinopathy (DR) is one of the leading causes of blindness globally. Earlier detection and timely treatment of DR are desirable to reduce the incidence and progression of vision loss. Currently, deep learning (DL) approaches have offered better performance in detecting DR from retinal fundus images. We, therefore, performed a systematic review with a meta-analysis of relevant studies to quantify the performance of DL algorithms for detecting DR.
A systematic literature search on EMBASE, PubMed, Google Scholar, Scopus was performed between January 1, 2000, and March 31, 2019. The search strategy was based on the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guidelines, and DL-based study design was mandatory for articles inclusion. Two independent authors screened abstracts and titles against inclusion and exclusion criteria. Data were extracted by two authors independently using a standard form and the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool was used for the risk of bias and applicability assessment.
Twenty-three studies were included in the systematic review; 20 studies met inclusion criteria for the meta-analysis. The pooled area under the receiving operating curve (AUROC) of DR was 0.97 (95%CI: 0.95-0.98), sensitivity was 0.83 (95%CI: 0.83-0.83), and specificity was 0.92 (95%CI: 0.92-0.92). The positive- and negative-likelihood ratio were 14.11 (95%CI: 9.91-20.07), and 0.10 (95%CI: 0.07-0.16), respectively. Moreover, the diagnostic odds ratio for DL models was 136.83 (95%CI: 79.03-236.93). All the studies provided a DR-grading scale, a human grader (e.g. trained caregivers, ophthalmologists) as a reference standard.
The findings of our study showed that DL algorithms had high sensitivity and specificity for detecting referable DR from retinal fundus photographs. Applying a DL-based automated tool of assessing DR from color fundus images could provide an alternative solution to reduce misdiagnosis and improve workflow. A DL-based automated tool offers substantial benefits to reduce screening costs, accessibility to healthcare and ameliorate earlier treatments.
糖尿病视网膜病变（DR）是全球致盲的主要原因之一。早期发现和及时治疗 DR 是降低视力丧失发生率和进展的理想方法。目前，深度学习（DL）方法在从眼底图像中检测 DR 方面表现出更好的性能。因此，我们进行了一项系统评价和荟萃分析，以量化 DL 算法检测 DR 的性能。
在 2000 年 1 月 1 日至 2019 年 3 月 31 日期间，对 EMBASE、PubMed、Google Scholar 和 Scopus 进行了系统的文献检索。搜索策略基于系统评价和荟萃分析的首选报告项目（PRISMA）报告指南，并且文章纳入必须基于 DL 研究设计。两名独立的作者根据纳入和排除标准筛选摘要和标题。两名作者使用标准表格独立提取数据，并使用诊断准确性研究的质量评估（QUADAS-2）工具评估偏倚风险和适用性。
共有 23 项研究纳入系统评价，20 项研究符合荟萃分析纳入标准。DR 的汇总接收者操作特征曲线（AUROC）下面积为 0.97（95%CI：0.95-0.98），敏感度为 0.83（95%CI：0.83-0.83），特异度为 0.92（95%CI：0.92-0.92）。阳性和阴性似然比分别为 14.11（95%CI：9.91-20.07）和 0.10（95%CI：0.07-0.16）。此外，DL 模型的诊断优势比为 136.83（95%CI：79.03-236.93）。所有研究均提供了 DR 分级量表，以人类分级者（例如，经过培训的护理人员、眼科医生）作为参考标准。
我们的研究结果表明，DL 算法在从眼底照片中检测可转诊 DR 方面具有较高的敏感度和特异度。应用基于 DL 的自动工具从彩色眼底图像评估 DR 可能是减少误诊和改善工作流程的一种替代方法。基于 DL 的自动工具具有显著的优势，可以降低筛查成本、提高医疗保健的可及性并改善早期治疗。
