To describe and evaluate a fully automated diabetic retinopathy screening system, for early diagnosis and reducing the work of ophthalmologists and alleviate inter and intra observer variability.

A set of 536 retinal images was selected from a diabetic retinopathy telescreening database. These images are either optic disc center or macula center which were captured from right and left eyes of 134 patients. All of the images have been manually graded by three ophthalmologists from Khatam-Al-Anbia eye hospital of Mashhad, Iran, using a software tool provided for image annotation. A fully automated diabetic retinopathy screening system which is called DRSRobo was evaluated. First a pre-screening algorithm is used to assess the quality of retinal images. If the quality of image is not satisfactory, it is examined by an ophthalmologist and reacquired if necessary. Otherwise, the image is processed for diabetic retinopathy detection. The proposed system is capable of detecting normal and abnormal retinal structures such as vessels, optic disc, macula, microaneuryms and exudates.

Sensitivity and specificity of 91.55% and 58.97% have been respectively achieved to detect abnormal retinal images. The system was also evaluated by receiver operating characteristics (ROC) curve and achieved an AUC of 76.35 for the same dataset.

Diabetic retinopathy is the major cause of blindness in urban populations. High costs of examination and treatment combined with the shortage of ophthalmologists prevent patients from getting regular follow-up, especially in rural areas. This paper provides a fully automated diabetic retinopathy screening system with the ability of retinal image quality assessment. This system with high sensitivity can be used as a screening tool in the reading centers to reduce the workload of the ophthalmologists.