ContransGAN: Convolutional Neural Network Coupling Global Swin-Transformer Network for High-Resolution Quantitative...

Area of Science:

• * Biology and Life Science
• * Optical Microscopy
• * Image Analysis

Background:

• * Optical quantitative phase imaging (QPI) is crucial for high-contrast biological cell detection and analysis.
• * Traditional optical microscopy faces challenges in directly obtaining quantitative phase information.
• * Microscope parameter trade-offs limit resolution, field of view (FOV), and depth of field (DOF).

Purpose of the Study:

• * To introduce a novel semi-supervised deep learning hybrid network, ContransGAN, for high-quality QPI.
• * To enable QPI acquisition using traditional optical microscopes across various magnifications.
• * To overcome the inherent limitations of traditional microscopy techniques.

Main Methods:

• * Development of ContransGAN, a hybrid network combining convolutional operations and multi-headed self-attention mechanisms.
• * Utilization of a semi-supervised learning approach requiring minimal unpaired microscopic images for training.
• * Integration of Convolutional Neural Network (CNN) for local feature extraction and Swin-Transformer for global feature extraction.

Main Results:

• * ContransGAN successfully generates high-quality quantitative phase images comparable to Transport of Intensity Equation (TIE) methods.
• * The framework accurately reconstructs high-resolution (HR) phase images from low-resolution (LR) bright-field images.
• * Demonstrated suitability for microscopic images across different resolutions and FOVs using biological and abiotic specimens.

Conclusions:

• * The proposed deep learning algorithm, ContransGAN, effectively enhances quantitative phase imaging capabilities in traditional optical microscopy.
• * ContransGAN offers a versatile solution for obtaining accurate phase information from diverse microscopic imaging conditions.
• * This advancement facilitates improved cell detection and analysis through accessible QPI.