Logo image
Back
Denoising diffusions in latent space for retinal layer segmentation
Conference proceeding

Denoising diffusions in latent space for retinal layer segmentation

Xiaojie Chen, Fahim Ahmed Zaman, Milan Sonka and Xiaodong Wu
Vol.13925, pp.139252P-139252P-5
Progress in Biomedical Optics and Imaging
04/03/2026
DOI: 10.1117/12.3086193

View Online

Abstract

Accurate segmentation of retinal layers in optical coherence tomography (OCT) images is essential for the diagnosis and monitoring of retinal diseases. However, traditional manual annotation is labor-intensive and subject to inter-observer variability, limiting its scalability in large-scale or longitudinal clinical studies. While deep learning–based methods alleviate the need for manual annotation during deployment, rather than eliminating the need for expert annotations during training, the goal of automated segmentation is to enable efficient, reproducible, and uncertainty-aware inference at scale, which remains inadequately addressed by existing deep learning models that often suffer from high computational cost or lack principled uncertainty quantification. We propose a novel retinal layer segmentation framework based on latent-space denoising diffusion probabilistic models (DDPMs) that leverages compact vectorized surface representations. By applying the diffusion process on latent representations derived from surface layer vectors, our method significantly reduces computational overhead while maintaining high segmentation accuracy. The generative nature of the model further enables reliable uncertainty estimation through repeated sampling. Experimental results on the JHU HCMS dataset show that our approach achieves an average Mean Absolute Error (MAE) of 0.968 pixels, outperforming deterministic regression baseline and demonstrating strong potential for clinical decision support.
 OCT retinal layer segmentation latent diffusion models

Details

Metrics

9 Record Views
Logo image