Purchase this article with an account.
Pok Chien Tan, Carol YL Cheung, Ecosse Lamoureux, Wynne Hsu, Mong Li Lee, Tien Yin Wong; Cloud-based Imaging Program for Diabetic Retinopathy Screening and Monitoring. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):1430.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
We tested the performance of a cloud-based automated imaging program for diabetic retinopathy (DR) screening and monitoring as used in a Singapore screening program, with the aim to reduce the workload of the trained graders for DR screening in reading centre.
The Singapore Integrated Diabetic Retinopathy Screening Programme (SiDRP) is a tele-medicine screening program for DR based on assessment of retinal images of patients with diabetes seen at the primary care setting. Currently, the assessment is based on a standardized assessment of DR presence and severity by trained non-ophthalmologists assessors (graders). To improve the efficiency of the SiDRP screening, we developed and machine-trained a prototype imaging software (SELENA), which is a cloud-based tele-medicine platform that processes digital retinal images based on pattern recognition and region extraction algorithms. The SELENA system aims to classify diabetic patients into those that need and do not need further assessment by trained graders. In addition, bringing the program to a cloud platform can open it up to global use, without geographical boundaries and limitations. The performance of the SELENA system was tested and verified using an independent dataset with varying severity of DR (200 diabetic patients) from the SiDRP. Using the ophthalmologists’ clinical assessment as the reference standard, the sensitivity, specificity, positive and negative predictive values were calculated.
Of the 200 patients, 84 had no DR, 66 had minimal/mild DR, 50 had moderate or above DR. In the analysis by the SELENA system, 4 patients were excluded as they were classified as ungradable. 85 tested positive with DR, 96 falsely positive, 15 negative, and 0 falsely negative. The sensitivity and negative predictive value of the SELENA software, were both of 100%. However, the software shows low specificity of 13.5% and a positive predictive value of 47%.
The SELENA system shows high sensitivity and negative predictive values, which indicate that the software is sensitive in identifying patients with a positive DR signal and ruling out those with no DR, which serves its purpose as an effective screening tool. Future work will focus on further refining the SELENA software to improve its specificity prior to its large scale use in the community, which will ultimately reduce the grading time for DR screening.
This PDF is available to Subscribers Only