Context-aware dynamic filtering network for confocal laser endomicroscopy image denoising

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Confocal laser endomicroscopy (CLE) is a key technology for diagnosing gastrointestinal diseases.
• Image noise, stemming from the fiber bundle's physical structure, significantly degrades CLE image quality.
• Existing denoising techniques struggle with realistic noise patterns in CLE images.

Purpose of the Study:

• To develop advanced denoising methods for realistic noise in CLE images.
• To improve the diagnostic accuracy of gastrointestinal disease detection using CLE.
• To create a self-supervised denoising model that does not require clean reference images.

Main Methods:

• Constructed a realistic noise model specific to CLE data.
• Developed context-aware kernel estimation (CLENet) to handle non-uniform noise.
• Integrated a multi-scale dynamic fusion module into a U-shaped network (U-CLENet).

Main Results:

• Achieved comparable or superior performance to existing methods on CLE image datasets.
• Demonstrated low computational complexity for the proposed denoising network.
• Successfully removed multiplicative and additive white noise from CLE images.

Conclusions:

• The proposed U-CLENet effectively enhances the visual quality of CLE images.
• Improved image clarity aids in diagnosing various stages of gastrointestinal tumors.
• The method supports computer-assisted diagnosis, potentially reducing clinical misdiagnosis rates.