Deep Learning Prediction of Retinal Thickness from Near-Infrared Fundus Photography: Toward Decentralized Quantitative Assessment of Diabetic Macular Edema

Objective: To predict pixel-wise retinal thickness maps from near-infrared (NIR) fundus images using deep learning (DL), and to identify image features in NIR fundus photographs serving as surrogate markers of retinal thickness, with implications for decentralized diabetic macular edema (DME) screening, progression monitoring, and treatment assessment. Methods: A DL model based on a U-Net architecture was trained on paired NIR fundus and OCT images from 531 eyes across three groups: healthy controls, diabetic retinopathy (DR) without DME, and DME. Model performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), structural similarity index (SSIM), and center-involved DME (ci-DME) classification at a central subfield thickness threshold of 300 µm. Controlled image manipulation experiments, including spatial disruption of vascular patterns, relocation of hard exudates, and contrast enhancement, were performed to identify image-level features serving as surrogate markers of retinal thickness. Results: The model achieved an MAE of 30.41 ± 18.68 µm, RMSE of 36.14 ± 21.05 µm, and SSIM of 0.87 ± 0.04 across the macula, with consistent performance across ETDRS subfields. For ci-DME classification, it achieved an accuracy of 84.1%, sensitivity of 69.1%, and specificity of 88.7%. Interpretability analyses were performed as qualitative assessments to visualize image regions contributing to model predictions. These analyses highlighted retinal vascular structures, hard exudates, and local contrast variations as visual features observed in relation to model outputs. Conclusions: NIR fundus images contain sufficient structural information to support pixel-wise retinal thickness estimation, with vascular architecture, hard exudates, and local contrast variations identified as image features potentially associated with model predictions. These findings suggest that NIR-based deep learning approaches may have potential applications in the assessment of diabetic macular edema and warrant further prospective and external validation to determine their role in screening, triage support, longitudinal monitoring, and treatment-related assessment, particularly in decentralized and re-source-limited care environments.