Abstract
Purpose: :
To assess the performance of automated analysis of retinal photographs for retinal disease detection in diabetes.
Methods: :
A set of images from 5,386 patients screened in 2007 were obtained from a non-midryatic Diabetic Retinopathy Screening Programme of the Central Region of Portugal. Images were taken at local Health Centres and graded by an experienced ophthalmologist. Moreover, to verify if the system misclassifies the referral cases considered urgent, 116 urgent cases registered between 2001 and 2007 were also collected for processing.The automated grading system (RetmarkerSR) consisted on a software for Microaneuryms (MAs) earmarking, generating two possible outputs: disease / no disease. RetmarkerSR results were compared with the ones manually graded.
Results: :
The results obtained with RetmarkerSR showed 2,560 patients (48%) as having ''no disease'' and therefore tagged for next screening and 2,826 patients (52%) as having ''disease'' and requiring manual grading by an expert. Comparing these results with the grading performed by an experienced ophthalmologist it is observed that all eyes considered urgent referrals are included in the ''disease'' group. Furthermore, from the 457 patients referred ''As Soon As Possible'' (ASAP) (6 month) by the ophthalmologist, RetmarkerSR only missed 18 cases (4%).Assessment of the system performance shows that RetmarkerSR achieves a sensitivity of 96.1% (95% CI (confidence interval) [94.39, 97.89]) with a specificity of 51.7% (95% CI [50.27; 53.07]).From the additional 116 cases collected and considered urgent by the ophthalmologist from 2001 to 2007, the RetmarkerSR showed that it misclassified only one case, found to be a macular pucker without any signs of coexistent diabetic retinopathy disease.
Conclusions: :
Automated grading of diabetic retinopathy may safely reduce the burden of ''disease/no disease'' grading in diabetic retinopathy screening programmes.
Keywords: imaging/image analysis: clinical • retina • diabetic retinopathy