Department of Ophthalmology and Visual Sciences, University of Louisville School of Medicine, Louisville, Kentucky, USA.
Department of Bioengineering, University of Louisville Speed School of Engineering, Louisville, Kentucky, USA.
Am J Ophthalmol. 2020 Aug;216:201-206. doi: 10.1016/j.ajo.2020.01.016. Epub 2020 Jan 23.
To determine if combining clinical, demographic, and imaging data improves automated diagnosis of nonproliferative diabetic retinopathy (NPDR).
Cross-sectional imaging and machine learning study.
This was a retrospective study performed at a single academic medical center in the United States. Inclusion criteria were age >18 years and a diagnosis of diabetes mellitus (DM). Exclusion criteria were non-DR retinal disease and inability to image the macula. Optical coherence tomography (OCT) and OCT angiography (OCTA) were performed, and data on age, sex, hypertension, hyperlipidemia, and hemoglobin A1c were collected. Machine learning techniques were then applied. Multiple pathophysiologically important features were automatically extracted from each layer on OCT and each OCTA plexus and combined with clinical data in a random forest classifier to develop the system, whose results were compared to the clinical grading of NPDR, the gold standard.
A total of 111 patients with DM II were included in the study, 36 with DM without DR, 53 with mild NPDR, and 22 with moderate NPDR. When OCT images alone were analyzed by the system, accuracy of diagnosis was 76%, sensitivity 85%, specificity 87%, and area under the curve (AUC) was 0.78. When OCT and OCTA data together were analyzed, accuracy was 92%, sensitivity 95%, specificity 98%, and AUC 0.92. When all data modalities were combined, the system achieved an accuracy of 96%, sensitivity 100%, specificity 94%, and AUC 0.96.
Combining common clinical data points with OCT and OCTA data enhances the power of computer-aided diagnosis of NPDR.
确定是否结合临床、人口统计学和影像学数据可提高非增殖性糖尿病视网膜病变（NPDR）的自动诊断。
横断面影像学和机器学习研究。
这是在美国一家学术医疗中心进行的回顾性研究。纳入标准为年龄＞18 岁且患有糖尿病（DM）。排除标准为非 DR 视网膜疾病和无法对黄斑成像。进行光学相干断层扫描（OCT）和 OCT 血管造影（OCTA），并收集年龄、性别、高血压、高血脂和糖化血红蛋白 A1c 等数据。然后应用机器学习技术。从 OCT 的每个层和每个 OCTA 丛中自动提取多个病理生理学上重要的特征，并将其与临床数据结合在随机森林分类器中，以开发系统，然后将其结果与 NPDR 的临床分级（金标准）进行比较。
共纳入 111 例 2 型糖尿病患者，其中 36 例无 DR，53 例轻度 NPDR，22 例中度 NPDR。当仅对 OCT 图像进行系统分析时，诊断的准确率为 76%，灵敏度为 85%，特异性为 87%，曲线下面积（AUC）为 0.78。当同时分析 OCT 和 OCTA 数据时，准确率为 92%，灵敏度为 95%，特异性为 98%，AUC 为 0.92。当结合所有数据模态时，系统的准确率为 96%，灵敏度为 100%，特异性为 94%，AUC 为 0.96。
将常见的临床数据点与 OCT 和 OCTA 数据相结合可增强计算机辅助 NPDR 诊断的能力。
