Untrained, physics-informed neural networks for structured illumination microscopy.

Area of Science:

• Microscopy
• Biophysics
• Computational Imaging

Background:

• Structured illumination microscopy (SIM) offers super-resolution imaging but traditional linear reconstruction methods can introduce artifacts and are limited by illumination patterns.
• Deep neural networks show promise for SIM reconstruction but typically require extensive, experimentally challenging training datasets.

Purpose of the Study:

• To develop a novel deep learning approach for SIM image reconstruction that eliminates the need for experimental training data.
• To create a versatile reconstruction method applicable to various SIM illumination patterns.

Main Methods:

• A physics-informed neural network (PINN) was combined with the forward model of the structured illumination process.
• The PINN was optimized using only diffraction-limited sub-images from a single dataset, incorporating illumination patterns into the loss function.
• The method was validated using both simulated and experimental SIM data.

Main Results:

• The PINN successfully reconstructed sub-diffraction images without requiring a dedicated training set.
• The method demonstrated versatility, adapting to different SIM illumination patterns by modifying the loss function.
• Achieved resolution improvements consistent with theoretical super-resolution limits.

Conclusions:

• Physics-informed neural networks offer a powerful, data-efficient alternative for structured illumination microscopy reconstruction.
• This approach overcomes limitations of traditional methods and data-hungry deep learning techniques.
• The PINN framework enables robust super-resolution imaging across diverse SIM configurations.