Self-Supervised Joint Learning for pCLE Image Denoising.

Area of Science:

• Medical Imaging
• Deep Learning
• Optical Engineering

Background:

• Probe-based confocal laser endoscopy (pCLE) is crucial for disease diagnosis but suffers from image artifacts like hexagonal patterns.
• Deep learning shows potential for image denoising, but training requires costly clean-noisy image pairs.
• Existing self-supervised denoising methods have limitations in complex imaging scenarios.

Purpose of the Study:

• To develop an innovative self-supervised denoising method for pCLE images.
• To address the challenge of acquiring paired training data for deep learning denoising models.
• To enhance image quality for improved diagnostic accuracy in pCLE.

Main Methods:

• Proposed a collaborative, jointly trained framework integrating noise prediction, image quality assessment, and denoising networks.
• Employed a self-supervised learning approach to eliminate the need for clean-noisy image pairs.
• Validated the method on both pCLE and fluorescence microscopy datasets.

Main Results:

• The proposed self-supervised method significantly improved image quality compared to existing approaches.
• Achieved superior denoising performance on pCLE images, reducing hexagonal pattern artifacts.
• Demonstrated effectiveness on fluorescence microscopy images, indicating broader applicability.

Conclusions:

• The novel self-supervised denoising technique enhances pCLE image quality for disease diagnosis.
• The synergistic integration of multiple networks offers a robust solution for artifact reduction.
• This method surpasses previous self-supervised techniques in both pCLE and fluorescence microscopy applications.