Automated Identification of Different Severity Levels of Diabetic Retinopathy Using a Handheld Fundus Camera and Single-Image Protocol

Abstract

PurposeTo evaluate the performance of artificial intelligence (AI) systems embedded in a mobile, handheld retinal camera, with a single retinal image protocol, in detecting both diabetic retinopathy (DR) and more-than-mild diabetic retinopathy (mtmDR). DesignMulticenter cross-sectional diagnostic study, conducted at three diabetes care and eye care facilities. ParticipantsA total of 327 individuals with diabetes mellitus (Type 1 or Type 2) underwent a retinal imaging protocol enabling expert reading and automated analysis. MethodsParticipants underwent fundus photographs using a portable retinal camera (Phelcom Eyer). The captured images were automatically analysed by deep learning algorithms RAS (retinal alteration score) and DRAS (diabetic retinopathy alteration score), consisting of convolutional neural networks trained on EyePACS datasets and fine-tuned using datasets of portable device fundus images. The ground truth was the classification of DR corresponding to adjudicated expert reading, performed by three certified ophthalmologists. Main Outcome MeasuresPrimary outcome measures included the sensitivity and specificity of the AI system in detecting DR and/or mtmDR using a single-field, macula-centered fundus photograph for each eye, compared to a rigorous clinical reference standard comprising reading center grading of 2-field imaging protocol using the ICDR severity scale. ResultsOf 327 analysed patients (age 57.0 + 16.8 years, diabetes duration 16.3 + 9.7 years), 307 completed the study protocol. Sensitivity and specificity of the AI system were high in detecting any DR with DRAS (sensitivity: 90.48%; 95% CI, 84.99% – 94.46% and specificity 90.65%; 95% CI, 84.54%-94.93%) and mtmDR with the combination of RAS and DRAS (sensitivity: 90.23%; 95% CI, 83.87%-94.69% and specificity: 85.06%; 95% CI, 78.88%-90.00%). The area under the ROC curve was 0.94 for any DR and 0.89 for mtmDR. ConclusionsThis study showed a high accuracy for the detection of DR in different levels of severity with a single retinal photo per eye in an all-in-one solution, composed of a portable retinal camera powered by AI. Such a strategy holds great potential for increasing coverage rates of screening programs, contributing to tackle avoidable blindness.