Simulation-driven learning: a deep learning approach for image scanning microscopy via physical imaging models.

Area of Science:

• Microscopy
• Deep Learning
• Image Reconstruction

Background:

• Deep learning is effective for fluorescence microscopy image reconstruction.
• Current methods require large datasets and extensive preprocessing for aliasing issues.

Purpose of the Study:

• To develop an improved generative adversarial network (GAN) for image scanning microscopy (ISM).
• To enable training with simulation data, enhancing network generalizability and reducing experimental burden.

Main Methods:

• Utilized physical imaging models to generate synthetic image pairs from simulation data.
• Trained a GAN using these simulated datasets, bypassing the need for real ISM images and alignment preprocessing.

Main Results:

• The simulation data-driven GAN improved the imaging quality of conventional microscopic images.
• The method successfully reduced the cost and complexity of experimental data acquisition.

Conclusions:

• The developed method offers an effective strategy for deep learning-based microscopy image reconstruction.
• This approach provides a pathway to optimize network generalizability using simulated data.