Abstract
Abstract
Endoscopic imaging faces challenges from complex anatomical structures, limited illumination angles, and variable environmental factors, which lead to inconsistent exposure and degrade image quality and diagnostic accuracy. To address this issue, we propose FSADiff, a Fourier spatial attention guided diffusion model that integrates global frequency modeling in the Fourier domain and spatial additive attention during the inverse diffusion process to jointly address the problem of inconsistent exposure. Specifically, Fourier transform computes global correlations in the frequency domain through element-wise multiplication, enabling effective capture of overall exposure deviations. An additive attention branch then adaptively modulates the frequency-domain results in the spatial domain to suppress local degradations. In addition, we introduce a dynamic noise embedding strategy that leverages a knowledge-aware network to incorporate temporal noise information into both the denoising network and the color corrector model, thereby improving image restoration performance. We evaluate FSADiff on public datasets, Endo4IE and Endovis17, as well as a proprietary multicenter clinical nasopharyngeal dataset. FSADiff achieved superior results, yielding a Peak Signal-to-Noise Ratio of 29.00 on Endo4IE and 33.27 on Endovis17 (all 6.5+ improvement over state-of-the-art). On the clinical nasopharyngeal dataset, FSADiff achieved a Blind / Referenceless Image Spatial Quality Evaluator score of 39.17 (7.02 improvement over state-of-the-art). Further evaluation on image subsets from three hospitals demonstrated significant improvements in both overall and individual quality metrics (
p
< 0.05).