Retinal thickness is one of the most important structural biomarkers in ophthalmology. Swelling of the retina, known as edema, is a hallmark of diabetic macular edema, one of the leading causes of vision loss in working-age adults, and thickness changes also track the progression of glaucoma, age-related macular degeneration, and other conditions. Clinicians measure it using optical coherence tomography, or OCT, which bounces near-infrared light into the eye to build depth-resolved cross-sectional images of the retinal layers. From those scans, software computes a thickness map centered on the fovea, the pit of sharpest vision. The problem is that OCT machines are costly, require trained operators, and remain unavailable in many rural clinics, screening programs, and low-resource health systems. Color fundus photography, by contrast, is fast, cheap, and widely deployed — but it captures only a two-dimensional surface view, with no direct information about depth or thickness.

The research team, led by Maryam Yahyaie and Reza AghaeiZadeh Zoroofi of the University of Tehran’s School of Electrical and Computer Engineering, together with ophthalmologists Alireza Ramezani and Zhale Rajavi of Shahid Beheshti University of Medical Sciences, asked a deceptively simple question: if a machine learns the statistical relationship between the surface appearance of the retina and its underlying thickness, could it infer the missing dimension? Their answer is a generative artificial intelligence framework they call the Anatomically Conditional Latent Diffusion Model, or AC-LDM, trained on paired datasets in which each fundus photograph is matched to the thickness map derived from the same patient’s OCT scan.

The technical heart of the system is a diffusion model, the same family of generative architectures that has transformed image synthesis in recent years. Diffusion models learn to create data by reversing a gradual noising process: during training, the model observes images corrupted by increasing amounts of random noise and learns to denoise them step by step. At generation time, starting from pure noise, it iteratively refines a random field into a coherent image that matches the conditions it is given — in this case, the fundus photograph. Rather than operating on raw pixels, which is computationally expensive, the framework works in a compressed latent space. A variational autoencoder first learns to squeeze each retinal thickness map into a compact perceptual representation that preserves essential structure while discarding irrelevant detail, and the diffusion process then unfolds within that efficient latent space.

What distinguishes AC-LDM from a generic image-to-image translator is its anatomical guidance. The researchers observed that a standard cross-attention module — the mechanism that lets a diffusion model consult the fundus photograph while generating the thickness map — treats all spatial locations equally, which can blur or distort modality-specific retinal structure. Their solution is a spatially weighted cross-attention module, or SW-CA, which injects anatomically informed spatial weights into the attention computation. Regions of the fundus image that carry more reliable information about retinal topography are given greater influence during generation, while less informative areas are down-weighted. The design draws on attention-based conditioning strategies developed in modern vision architectures, but adapts them to the specific geometry and physiology of the retina, encouraging the model to respect the biological correspondence between what a fundus camera sees and what OCT measures.

Evaluating generative medical images is notoriously difficult, because a plausible-looking output is not necessarily an accurate one. The team therefore benchmarked AC-LDM against representative convolutional, generative adversarial, and transformer-based synthesis approaches using a battery of established metrics. Peak signal-to-noise ratio, or PSNR, quantifies pixel-level fidelity between the synthesized and true thickness maps; the structural similarity index, or SSIM, captures preservation of local structure such as the foveal depression and perifoveal rings. Two perceptual and distributional measures round out the picture: LPIPS, the learned perceptual image patch similarity, which correlates with human judgments of visual difference, and FID, the Fréchet Inception distance, which compares the statistical distribution of generated images against real ones.

The results were striking. AC-LDM achieved a PSNR of 30.88 decibels and an SSIM of 0.871, indicating both high pixel-level agreement and strong structural fidelity with the ground-truth thickness maps. On the perceptual side, it recorded an LPIPS of 0.105 and an FID of 29.5, both lower — and therefore better — than the baseline models it was compared against. Lower LPIPS means the synthesized maps look more like real ones to learned perceptual features, while the lower FID indicates that the overall population of generated maps is statistically closer to authentic retinal topography. Together, these numbers suggest that the anatomically guided diffusion approach does not merely produce attractive images but captures genuine structural information about the retina.

The implications extend well beyond a single laboratory benchmark. If thickness maps can be reliably synthesized from fundus photographs, screening programs for diabetic retinopathy — which already rely heavily on fundus photography in community and telemedicine settings — could gain a proxy measure of macular edema without deploying OCT hardware. The framework also opens a door for computational ophthalmology more broadly: large-scale fundus image archives, collected over decades and numbering in the millions, could potentially be retrofitted with synthesized thickness information, enabling retrospective studies of retinal structure at a scale OCT has never achieved. Cross-modal synthesis of this kind may likewise support data augmentation, privacy-preserving data sharing, and the training of downstream diagnostic models.

The authors are careful, however, to frame the work as a research advance rather than a clinical tool. Their own conclusion notes that clinical validation with thickness-specific error metrics — such as mean absolute error, root mean squared error, and agreement measures for central foveal thickness under the standard ETDRS grid — remains necessary before any translational use. Image-fidelity and perceptual scores, while encouraging, do not by themselves prove that synthesized thickness values are accurate enough to guide treatment decisions such as initiating anti-VEGF injections for macular edema. The study was approved by the Research Ethics Committee of Shahid Beheshti University of Medical Sciences, conducted under the Declaration of Helsinki, and used fully anonymized retrospective imaging data hashed to comply with HIPAA and GDPR requirements.

Still, the trajectory is clear and compelling. Diffusion models have already conquered natural image generation, and this study demonstrates how domain knowledge — in this case, the anatomy of the retina — can be woven into their attention mechanisms to solve a genuinely medical problem. The work, which received no specific funding and is published open access, points toward a future in which the humble fundus camera, already ubiquitous from Tehran to rural screening vans, becomes a window not just onto the surface of the retina but into its hidden third dimension. For the hundreds of millions of people at risk of retinal disease worldwide, many of whom live where OCT machines may never reach, that window could make the difference between blindness detected too late and sight preserved in time.

Subject of Research: Cross-modal synthesis of retinal thickness maps from color fundus photographs using an anatomically guided latent diffusion model

Article Title: Latent diffusion with anatomical guidance for retinal thickness map synthesis

Article References: Yahyaie, M., Zoroofi, R. A., Ramezani, A., & Rajavi, Z. (2026). Latent diffusion with anatomical guidance for retinal thickness map synthesis. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02792-4

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02792-4

Keywords: artificial intelligence, diffusion models, retinal thickness map, color fundus photography, optical coherence tomography, medical imaging, ophthalmology, deep learning, cross-modal synthesis, macular edema, diabetic retinopathy, computational ophthalmology