Hybrid diffusion model for OCT-angiography vessel segmentation with denoising enhancement

Abstract

Optical Coherence Tomography Angiography (OCTA) technology provides detailed visualization of the retinal vascular system, where accurate vessel segmentation is crucial for diagnosing vision-related diseases. However, the 3D volume data, affected by inherent modality constraints, contains artifacts and noise, complicating precise vessel extraction in down-sampled images. To address these challenges, this study introduces a hybrid model architecture. The proposed method leverages a diffusion model to learn the underlying noise distribution and regulate the denoising process by controlling the time step. This facilitates noise suppression, vascular structure restoration, and enhanced vessel-background contrast. Furthermore, we design a lightweight segmentation discriminator that utilizes denoised images as conditional inputs. By leveraging wavelet convolution, the discriminator extracts both high- and low-frequency features, enhancing texture representation and detail preservation. This ultimately contributes to more precise vessel segmentation. The diffusion model and segmentation discriminator are incorporated into a unified end-to-end network framework. Extensive experiments on the OCTA-500 and ROSE-1 datasets validate the superiority of our method over state-of-the-art approaches in vessel segmentation.