Deep-learning model achieved high diagnostic performance for diabetic macular edema detection using optical coherence tomography (OCT) scans
1. Nguyen and colleagues developed and evaluated a deep-learning model for detecting diabetic macular edema using 3D OCT scans.
2. The model achieved accuracy, sensitivity, and specificity above 90% in a real-world screening program and flagged uncertain cases for specialist review.
Evidence Rating Level: 2 (Good)
Study Rundown: Diabetic macular edema (DME) is the primary cause of vision loss among individuals with diabetes, and early detection of DME is crucial for preventing and delaying visual impairment. Nguyen and colleagues developed a privacy-preserving deep-learning model to detect DME and distinguish center-involved DME (CI-DME) from non-CI-DME using 3D OCT scans. The model was trained through federated learning on multicenter data without transferring patient images between institutions. The model was then prospectively evaluated in a diabetic retinopathy screening program in Vietnam. The main performance metrics were accuracy, sensitivity, specificity, and comparisons with ophthalmologists. The study found that the model achieved an accuracy of 93.70%, a sensitivity of 91.78%, and a specificity of 93.06%. Diagnostic performance was lower for distinguishing CI-DME from non-CI-DME. The model also identified 64 uncertain cases for ophthalmologist review and performed similarly to human experts. This study demonstrated that a deep-learning model can be used to expand DME screening in lower-resourced settings.
Click here to read the study in NEJM AI
Relevant Reading: Real-world performance of an AI system for diabetic retinopathy screening
In-Depth [prospective cohort]: The model was developed and evaluated using 8,031 OCT volumes from 1,958 adults across five centers in Hong Kong, Singapore, and the United States. Three Hong Kong datasets were used for federated training, while Singaporean and American datasets were used for external retrospective testing. The model first detected DME and then classified positive scans as CI-DME or non-CI-DME. Federated averaging enabled collaborative training without sharing patient images. For prospective testing, 1,473 OCT volumes from 753 consecutively recruited participants in a Vietnamese diabetic retinopathy screening program were analyzed in real time. OCT scans were classified as positive, negative, or uncertain; uncertain cases were referred to ophthalmologists. Performance metrics included area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, and specificity. The model achieved 93.70% accuracy (95% confidence interval [CI], 91.24-94.01%), 91.78% sensitivity (95% CI, 86.84-94.36%), and 93.06% specificity (95% CI, 91.53-94.49%) for DME detection. CI-DME classification achieved 83.75% accuracy (95% CI, 78.17-88.83%), 85.61% sensitivity (95% CI, 79.56-91.17%), and 79.31% specificity (95% CI, 68.75-89.09%). A total of 64 cases were identified as uncertain and were referred to ophthalmologists. This study was limited by the use of a single commercial OCT device, imperfect specificity that could increase referrals, and uncertainty regarding the optimal management of borderline cases. Overall, this study demonstrated that DME can be effectively screened using a deep-learning model to expand access in lower-resourced settings.
Image: PD
©2026 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
