Artificial intelligence algorithms for the diagnosis of signs of diabetic retinopathy, diabetic macular edema, age-related macular degeneration, vitreomacular interface abnormalities

Abstract

Purpose. Development of artificial intelligence (AI) algorithms for diagnosing of diabetic retinopathy (DR), diabetic macular edema (DME), age-related macular degeneration (AMD), vitreomacular interface abnormalities (VMA) through the analysis of OCT scans and fundus images. Material and methods. Fundus images of patients with DR and DME, OCT scans of patients with DME, AMD and VMA were used as training and validation databases. The volume of training databases was 3600 fundus images and 10 000 OCT scans, the volume of validation databases was 400 fundus images and 1000 OCT scans. For fundus images analysis algorithms accuracy, sensitivity, specificity, AUROC were calculated for the following structures: microaneurysms, intraretinal hemorrhages, hard exudates, soft exudates, retinal and optic disc neovascularization, preretinal hemorrhages, epiretinal fibrosis, laser coagulates. For OCT scan analysis algorithms, these metrics were calculated for the features: intraretinal cysts, subretinal fluid, pigment epithelium detachment, subretinal hyperreflective material, drusen, epiretinal membrane, full thickness macular hole, lamellar macular hole, vitreomacular traction. Results. For fundus images analysis algorithms, accuracy exceeded 93% for all features except soft exudates (88.3%) and neovascularization (88.0%), sensitivity exceeded 90% for all features except neovascularization (80.2%) and epiretinal fibrosis (72.5%), specificity exceeded 91% for all features except microaneurysms (80.5%), hard exudates (83.5%) and soft exudates (88.7%), AUROC exceeded 0.90 for all signs except epiretinal fibrosis (0.88), neovascularization (0.87), preretinal hemorrhages (0.89). For OCT analysis algorithms, accuracy exceeded 93% for all features, sensitivity exceeded 90% for all features except lamellar macular hole (87.22%), specificity exceeded 93% for all features, AUROC exceeded 0.93 for all features. Conclusion. An algorithm for high precision segmentation of pathological signs has been developed. Based on these AI algorithms, the Retina.AI ophthalmological platform was developed, which allows automated analysis of OCT scans and fundus images and diagnosing of DR, DME, AMD and VMA. The platform is available for testing at https://www.screenretina.com/ Keywords: artificial intelligence, ophthalmic screening, diabetic retinopathy, diabetic macular edema, age-related macular degeneration, vitreomacular interface abnormalities