Purpose
Automated detection program has the potential to reduce the effort of image reading by raters. The purpose of this study is to evaluate the performance of an automated detection program for diabetic retinopathy (DR) and referable DR (RDR) from digital retinal photographs.
Methods
Color retinal images of both DR and normal subjects were collected from Zhongshan Ophthalmic Center. The 45-degree photographs centered on the macular were chose for the present analysis. Image gradability was assessed according to the DR screening scheme of the United Kingdom National Screening Committee. Retinal color photographs were graded independently for retinopathy severity according to the International Clinical Diabetic Retinopathy and Diabetic Macular Edema Disease Severity Scales by two trained graders and adjudicated by a retinal specialist. Gradable images were then analyzed by the automated assessment program. Referable DR was defined as severer than mild non-proliferative DR with or without macular edema.
Results
A total of 289 gradable images were assessed by the automated program. The sensitivity and specificity for DR detection was 91.6% (95%CI, 85.8% to 95.6%) and 59.6% (95%CI, 51.2% to 67.6%), respectively. In terms of RDR, the sensitivity of the program raised to 93.5% (95%CI, 88.1% to 97.0%), and specificity was 60.0% (95%CI, 51.7% to 67.9%). The program was failed to detect retinopathies in 9 of 139 DR cases (false-negative ratio: 3.10%). All the nine images met the quality criteria for manual grading but were relatively inadequate for autodetection.
Conclusions
Our automated detection program has high sensitivity and specificity to detect DR and RDR. It can be served as an assistant tool safely in DR screening projects, potentially reducing the workload of graders on image reading.