Purchase this article with an account.
Kang Wang, Chaitra Jayadev, Muneeswar gupta Nittala, Chaithanya Ramachandra, Malavika Bhaskaranand, Sandeep Bhat, Kaushal Solanki, Srinivas R Sadda; Automated Detection of Diabetic Retinopathy Lesions on Ultrawidefield Pseudocolor Images. Invest. Ophthalmol. Vis. Sci. 2016;57(12):6366. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Peripheral lesions have been shown to be of prognostic significance in eyes with diabetic retinopathy (DR), raising interest in the use of ultrawidefield (UWF) images for screening and staging DR. In this study, we sought to determine the sensitivity and specificity of an automated algorithm for detecting referral-warranted DR in Optos UWF pseudocolor images.
383 subjects (739 eyes) with diabetes, totaling 1661 UWF pseudocolor images from Optos Daytona device sent from Narayana Nethtralaya (Bangalore, India) were analyzed at the Doheny Image Reading Center (DIRC) by certified DR graders. Moderate Non-proliferative DR (NPDR) or higher on the 5-level International Clinical Diabetic Retinopathy (ICDR) severity scale was considered as grounds for "referral". Fifty percent of images (“training set”) were used to train the classifiers in the automated algorithm (modified version of the EyeArt automated DR software that has been validated for standard flash color images). The other half of the dataset (“test set”) was used to assess the system performance. The software automatically detected DR lesions (hemorrhages, microaneurysms, lipid exudates, cotton wool spots) using the previously trained classifiers, and classified each image in the test set as “referral warranted” or “not warranted.” Sensitivity and specificity of the referral classification were computed.
The distribution of DR among the study eyes was: No DR=393, mild NPDR=7, Moderate NPDR=139, Severe NPDR=44 and PDR=38. 118 eyes were not graded due to poor image quality, lash artifacts. An ungraded UWF pseudocolor image and the same image following automated DR lesion detection is shown in Figure 1. For detection of referral warranted retinopathy, the automated algorithm achieved 90% sensitivity (95% CI 88.1-91.8) with 65.4% specificity (95% CI 46.4–84.0).The AUROC (Area under Receiver Operating Curve) of the algorithm was 0.87 (95% CI 0.82-0.92).
DR lesions could be detected from Optos UWF pseudocolor images using an automated algorithm, and the images could be classified as referral warranted diabetic retinopathy with a high degree of sensitivity and moderate specificity. Automated analysis of UWF images could be of value in DR screening programs, and for producing a more complete and accurate staging of the disease. Larger, prospective studies are required.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only