Diabetic Retinopathy Screening Using an Expert System with an Operator-Free Nidek Orion Fundus Camera

Worldwide, 240 million people have diabetes with 50% unaware of their condition. An estimated 2.5 million have diabetic retinopathy (DR), which is the leading cause of adult blindness. Fundus photography reading centers are overwhelmed by the growing number of DR cases to review. This study describes an automated artificial intelligence (AI) system that screens for DR. Subjects were recruited from the patient population at a medical university diabetes clinic. Dilated eyes (359 non- DR & 18 DR) and undilated eyes (130 non-DR & 11 DR) were used. Non-DR included normal and diabetic subjects with normal retinas or non-visually threatening (VT) disease. DR included only cases of VT disease. Cases in which retinal imaging was impossible (e.g., severe cataracts) were excluded from the study. A Nidek Orion non-mydriatic automated fundus camera (Nidek Technologies, srl, Padova, Italy) recorded 5 overlapping fundus fields that were combined into a single 45° mosaic image. Canon fundus photographs were also acquired at the same clinic visit using a 3-field modified ETDRS pattern. The Canon images were later screened by a human expert at a major fundus photography reading center to serve as a comparative gold standard. The Orion mosaic images were processed for contrast enhancement and passed through an analysis routine that extracted the location, size, and shape of bright and dark blobs. This information was passed to an Expert System that determined if DR was present and if it was VT or non-VT based on location and size. Only VT results were flagged for DR output. Results of the Expert System were then compared to the gold standard using the same criteria. Dilated eyes had 78.6% accuracy, 94.4% sensitivity, and 78.8% specificity, while undilated eyes had 98.2% accuracy, 90.9% sensitivity, and 98.7% specificity in screening. Additional tuning is needed to reduce VT-like artifacts seen with pupil dilation. This Expert System is an effective step toward mass screening of undilated subjects with a risk for VT DR. The AI software can be contained within the Orion, providing rapid analysis and avoiding concerns about data integrity and patient privacy that are associated with telemedicine methods. The expert system distinguishes the referable VT DR from non-VT cases, thus saving time and improving the management of this blinding eye disease.