Unpaired data training enables super-resolution confocal microscopy from low-resolution acquisitions

Area of Science:

• Microscopy
• Deep Learning
• Image Processing

Background:

• Super-resolution imaging enhances microscopic image resolution beyond the diffraction limit.
• Supervised deep learning requires large, paired datasets, posing practical challenges.
• Cycle-consistent generative adversarial networks (cycleGANs) offer an unsupervised alternative for image-to-image translation using unpaired data.

Purpose of the Study:

• To develop and evaluate a cycleGAN framework for unsupervised super-resolution in confocal microscopy.
• To increase the lateral resolution of confocal images without relying on paired low- and high-resolution training data.

Main Methods:

• A cycleGAN model was trained using unpaired low- and high-resolution confocal images of human glioblastoma cells.
• Performance was assessed using metrics like background standard deviation, peak-to-noise ratio, and frequency content.
• Image fidelity and resolution improvement were evaluated using a paired dataset.

Main Results:

• The cycleGAN model successfully increased the lateral resolution limit in confocal microscopy.
• The model demonstrated superior performance in image fidelity and resolution enhancement compared to other methods.
• Quantitative metrics confirmed the effectiveness of the unsupervised approach.

Conclusions:

• CycleGAN models are effective for super-resolution microscopic imaging without the need for paired training data.
• This approach enables the transformation of low-resolution microscopes into cost-effective super-resolution instruments.
• Unsupervised deep learning offers a promising solution for advancing microscopic imaging capabilities.