Deep Learning Prediction of Retinal Thickness Map from Two-Dimensional Fundus Photography

Retinal thickness measurement is important for diagnosing and managing diabetic macular edema (DME). Optical coherence tomography (OCT) is the clinical standard for this purpose. However, OCT is expensive and not always available in primary care or remote settings. This study demonstrates the feasibility of using deep learning to predict retinal thickness maps directly from two-dimensional fundus photographs. Paired fundus and OCT images were used to train and evaluate the model. The predicted thickness maps showed strong agreement with OCT-derived maps and accurately identified eyes with center-involved DME (ci-DME). Further analysis showed that model performance was influenced by anatomical and pathological cues, such as vascular patterns and hard exudates. These findings indicate that fundus images contain sufficient information to estimate retinal thickness patterns across the macula. The proposed approach offers an affordable and practical solution for decentralized eye care and may help reduce unnecessary OCT referrals while improving access to DME screening and monitoring.