Core Concepts
Deep learning GANs enhance biological image quality across microscopy systems.
Abstract
Introduction
High-quality microscopy is crucial in biological sciences.
Limitations include cost and imaging throughput.
Basic wide-field fluorescence microscopy is more accessible.
Methods
GAN architecture with discriminator and generator.
Training on paired images from different microscopes.
U-NET for generator and CNN for discriminator.
Results and Discussion
GAN outperforms deconvolution in image quality.
Network generalizes HQ features to LQ images effectively.
Metrics show high similarity between generated and ground truth images.
Conclusions
GAN model robust for image quality enhancement.
Potential for multi-institute microscopy cooperative network.
Stats
For our best model, median values are 6·10^-4 for MSE, 0.9413 for SSIM, and 31.87 for PSNR.
Median values for LQ vs. HQ ground truth are 0.0071 for MSE, 0.8304 for SSIM, and 21.48 for PSNR.
Quotes
"Our model proves that transfer between microscopy systems is possible."
"Results show a significant increase in SSIM and PSNR values."