Automated detection of diabetic retinopathy: Results of a screening study

Background: This study evaluated the operating characteristics of a reading software (RetinalyzeO System, Retinalyze A/S, Horsholm, Denmark) for automated prescreening of digital fundus images for diabetic retinopathy. Methods: Digital fundus images of patients with diabetes were retrospectively selected from the Bro Taf diabetic retinopathy screening program in Wales, UK in the period of 2002-2004, which has been superseded by the Diabetic Retinopathy Screening Service for Wales. A gold standard reference was defined by classifying each patient as having or not having diabetic retinopathy based on overall visual grading of the digitized images using the Bro Taf reading protocol. Automated grading was applied using automated red or bright lesion detection at varying detection sensitivities and adjusting for image quality. Operating characteristics included sensitivity, specificity, positive predictive values, and negative predictive values (PPV and NPV, respectively). Results: Automated analysis of four hundred fundus photographs of 192 eyes from 96 patients with diabetes was performed. The automated red lesion detection had a sensitivity of 82%, specificity of 75%, PPV of 41%, and NPV of 95%. Combined automated red and bright lesion detection yielded a sensitivity of 88%, specificity of 52%, PPV of 28%, and NPV of 95%. Performance of the combined red and bright lesion detection at elevated thresholds in images of good quality demonstrated a sensitivity of 93%, specificity of 78%, PPV of 46%, and NPV of 98%. Conclusions: Prescreening for diabetic retinopathy by automated detection of single fundus lesions seem to be achieved with minimal false negativity and can help to decrease the burden of manual diabetic retinopathy screening.