Advancing Diabetic Macular Edema Detection from 3D Optical Coherence Tomography Scans: Integrating Privacy-Preserving AI and Generalizability Techniques — A Prospective Validation in Vietnam

BackgroundDiabetic macular edema (DME) is the primary cause of irreversible vision loss among people with diabetes and can be accurately detected by using optical coherence tomography (OCT). We developed and validated a deep learning (DL) model to classify DME on OCT volumetric scans, enhanced by federated learning and advanced DL methods to safeguard patient privacy and improve model generalizability in analyzing unseen OCT scans. The performance and effectiveness of the DL model were then prospectively evaluated in a real-world diabetic retinopathy (DR) screening program in Vietnam.MethodsWe developed and externally tested a federated learning–based DL algorithm for detecting DME and further classifying center-involved DME (CI-DME) and non-CI-DME through three-dimensional OCT volumetric scans. The study used 8031 OCT volumes from 1958 participants with diabetes from Hong Kong, the United States, and Singapore. This DL model was prospectively tested with a novel test-time adaptation method in real time on 1473 OCT volumes from 753 participants with diabetes in a DR screening program in Vietnam. An uncertainty range including dual thresholds was newly introduced to improve the model's trustworthiness by flagging uncertain cases in real-world clinical application.ResultsIn the prospective study in Vietnam, the DL model showed accuracy of 93.70% (95% confidence interval [CI], 91.24 to 94.01%), sensitivity of 91.78% (95% CI, 86.84 to 94.36%), and specificity of 93.06% (95% CI, 91.53 to 94.49%) for detecting the presence of DME, and it showed accuracy of 83.75% (95% CI, 78.17 to 88.83%), sensitivity of 85.61% (95% CI, 79.56 to 91.17%), and specificity of 79.31% (95% CI, 68.75 to 89.09%) for differentiating CI-DME and non-CI DME. In addition, the model identified 64 cases as uncertain, indicating a need for re-evaluation by an ophthalmologist. The DL model and human experts achieved similar performance in identifying DME among individuals with diabetes.ConclusionsOur DL model was effective in detecting DME from three-dimensional OCT scans in a prospective, real-time clinical setting, suggesting that successful deployment of DL to improve DR screening programs in lower- and middle-income countries can be achieved. (Funded by the General Research Fund and others.)