Cycle-Consistent Diffusion Model With Vessel-Aware Attention for Endoscopic Image Translation

Abstract

Diffusion models have emerged as a powerful tool for image-to-image translation due to their high visual fidelity and stable training, but their application in medical imaging is limited by their tendency to degrade structural information during the forward diffusion. This work introduces a diffusion-based framework that integrates explicit anatomical guidance into the generation process. Specifically, vessel-aware attention is injected into the denoising network in a timestep-aware manner to preserve vascular and anatomical structures throughout the denoising trajectory. The proposed method is applied to unpaired translation from White Light (WL) to Narrow Band Imaging (NBI) in cystoscopy, demonstrating superior anatomical fidelity and visual realism over existing methods.