Generating 3D Optical Coherence Tomography from 2D Fundus Images via Diffusion Models

Training machine learning models with synthetic data effectively addresses data scarcity, particularly in domains where acquiring large-scale 3D datasets is costly. We present Fundus2OCT, the first framework to synthesize high-fidelity 3D optical coherence tomography (OCT) volumes from 2D fundus photographs using diffusion models. Developed on paired fundus-OCT data from the UK Biobank, Fundus2OCT leverages a two-stage latent diffusion process to generate anatomically coherent OCT volumes (32 B-scans per volume) conditioned on fundus inputs. Quantitative evaluations demonstrate superior performance over existing methods, with Fréchet Inception Distance (FID) and Fréchet Video Distance (FVD) scores of 12.3 and 58.7, respectively. In a clinical Turing test, two ophthalmologists achieved accuracies of 49.0–57.0% (near chance-level) in distinguishing synthetic from real OCTs. To validate clinical utility, we augmented four public fundus-based disease detection tasks (AMD, glaucoma, DR, DME) with synthetic OCT data, improving multimodal classification AUC by 4.2–8.6%. By bridging 2D fundus findings with 3D structural insights, Fundus2OCT advances multimodal retinal analysis, offering a scalable solution to enhance diagnostic accuracy and accessibility in ophthalmic care.