Fourier spatial attention guided diffusion model for optimizing exposure inconsistencies in endoscopic images.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Endoscopic imaging quality is often compromised by complex anatomy, poor lighting, and environmental variability, leading to inconsistent exposure.
• These factors degrade image quality and reduce diagnostic accuracy in clinical settings.

Purpose of the Study:

• To introduce FSADiff, a Fourier spatial attention guided diffusion model, designed to overcome inconsistent exposure issues in endoscopic imaging.
• To improve the overall quality and diagnostic utility of endoscopic images.

Main Methods:

• FSADiff integrates global frequency modeling in the Fourier domain and spatial additive attention during the inverse diffusion process.
• It utilizes Fourier transform for global exposure deviation capture and an additive attention branch for local degradation suppression.
• A dynamic noise embedding strategy incorporates temporal noise information using a knowledge-aware network.

Main Results:

• FSADiff demonstrated superior performance on public datasets (Endo4IE, Endovis17) and a clinical nasopharyngeal dataset.
• Achieved significant improvements in Peak Signal-to-Noise Ratio (PSNR) and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) scores, surpassing state-of-the-art methods.
• Clinical evaluation across three hospitals confirmed significant improvements in image quality metrics (p < 0.05).

Conclusions:

• FSADiff effectively addresses inconsistent exposure in endoscopic imaging through its unique Fourier and spatial attention mechanisms.
• The proposed model offers a substantial advancement in endoscopic image restoration, enhancing diagnostic accuracy.
• The findings suggest broad applicability and potential for clinical integration of FSADiff.