Cycle-consistent deep learning approach to coherent noise reduction in optical diffraction tomography.

Area of Science:

• Optical Imaging
• Machine Learning
• Image Processing

Background:

• Coherent noise is a significant challenge in 3D quantitative phase imaging, degrading image quality and hindering analysis.
• Existing denoising methods often struggle with specific noise types or require paired clean and noisy data.

Purpose of the Study:

• To develop and validate a deep neural network for effective coherent noise reduction in 3D quantitative phase imaging.
• To demonstrate the network's ability to handle unpaired image datasets, a novel approach for optical imaging problems.

Main Methods:

• A deep neural network, inspired by cycle generative adversarial networks, was trained to learn transformations between noisy and clean refractive index tomograms.
• The network was trained using unpaired datasets of clean and noisy images.
• Performance was evaluated through denoising experiments on various samples.

Main Results:

• The deep neural network successfully reduced coherent noise in 3D quantitative phase imaging.
• The method demonstrated strong performance and generalization capabilities across different samples.
• The technique was successfully applied to reduce temporal noise from focal drift in time-lapse imaging of biological cells.

Conclusions:

• The proposed deep neural network offers a powerful and versatile solution for coherent noise reduction in 3D quantitative phase imaging.
• This unpaired learning approach overcomes limitations of previous methods and enables denoising in challenging scenarios like time-lapse imaging.
• The technique provides a unique capability for reducing temporally changing noise, such as that caused by focal drift, which is not achievable with conventional optical methods.