Deep learning strategies for estimating retinal thickness from fundus images: a comparative study with multi-device data

Color fundus photographs (CFPs) are widely used for diabetic retinopathy (DR) screening due to accessibility. Optical coherence tomography (OCT) provides complementary depth-resolved retinal structure information but is less practical for large-scale screening, especially due to its cost. This study presents a deep learning framework based on a foundation model (FM) to estimate OCT-derived total retinal thickness (TRT) maps directly from CFPs, without preprocessing steps such as region registration. RETFound was used as the backbone FM, and the AI-READI dataset with paired CFPs and OCT scans from two devices (Topcon Maestro2 and Triton) was used for training and evaluation. Additional training data came from OCTA-500 and EyePACS. Three decoder training strategies for TRT estimation were compared: Individual, Combination, and Random Selection. Device-derived OCT TRT maps showed a significant inter-device thickness difference, with Maestro2 measuring on average more than 30 µm thicker than Triton. Individual and Combination models achieved low bias (relative difference about 5%) against device-specific references. The Random Selection model converged toward the average TRT distribution between Maestro2 and Triton, demonstrating robustness to device variability. The proposed FM-based framework bridges CFP and OCT by estimating retinal thickness from CFPs without OCT.