RepE: unsupervised representation learning for image enhancement in nonlinear optical microscopy.

Area of Science:

• Biomedical imaging
• Microscopy
• Machine learning

Background:

• Nonlinear optical microscopy, including second harmonic generation (SHG) and two-photon fluorescence (TPEF), is crucial for biological and medical imaging.
• Image noise is a significant challenge in these modalities, hindering accurate analysis, particularly in pathological samples like esophageal cancer tissue slides (ESCC).
• Existing denoising methods often require paired clean and noisy images or rely on specific statistical assumptions, limiting their applicability.

Purpose of the Study:

• To introduce RepE (representation and enhancement), an unsupervised learning method for denoising nonlinear optical microscopy images.
• To address the limitations of current denoising techniques by eliminating the need for paired data and statistical assumptions.
• To evaluate the performance of RepE on real-world SHG and TPEF images from esophageal cancer tissue slides.

Main Methods:

• RepE utilizes an encoder network to learn noise-free image representations.
• A reconstruction network is employed to generate denoised images from these learned representations.
• The method requires only a small number of training images and operates without restrictive statistical assumptions.

Main Results:

• RepE successfully denoises nonlinear optical microscopy images with various noise types.
• Comparative evaluations show that RepE outperforms existing denoising techniques on real-world SHG and TPEF images from ESCC slides based on image quality metrics.
• The method demonstrates robustness and effectiveness in practical applications.

Conclusions:

• RepE offers a practical and robust unsupervised learning solution for denoising nonlinear optical microscopy images.
• The method's ability to work without clean-noisy pairs and statistical assumptions makes it broadly applicable.
• RepE has the potential for extension to other nonlinear optical microscopy modalities, advancing image analysis in biomedical research.