Untrained Position-Encoded Multilayer Perceptron Network for Structured Illumination Microscopy Reconstruction

Area of Science:

• Microscopy
• Computational Imaging
• Biophysics

Background:

• Structured Illumination Microscopy (SIM) achieves super-resolution by encoding spatial details with patterned light.
• Conventional Fourier-based SIM reconstruction methods can produce artifacts, while deep learning approaches often require extensive training data and lack flexibility.
• Existing methods struggle with adaptability across different imaging setups and suboptimal imaging conditions.

Purpose of the Study:

• To develop a novel, data-efficient method for reconstructing super-resolution images from SIM data.
• To overcome limitations of traditional and deep learning-based reconstruction techniques in SIM.
• To enable robust and adaptable super-resolution imaging without the need for large training datasets.

Main Methods:

• Development of a Position Encoded Multi-Layer Perceptron (PEM) network utilizing implicit neural representations (INRs).
• Integration of a SIM forward model for iterative reconstruction optimization based on structural similarity loss.
• Encoding spatial coordinates using multi-frequency sinusoidal functions for detailed feature representation.

Main Results:

• PEM-SIM successfully reconstructs 2D and 3D super-resolution images with fewer input frames compared to conventional methods.
• The method demonstrates robustness across varying signal-to-noise ratios and performs comparably to standard algorithms on synthetic and experimental data.
• PEM-SIM effectively predicts missing axial planes in 3D SIM datasets.

Conclusions:

• PEM-SIM offers a data-efficient and adaptable solution for super-resolution imaging in microscopy.
• The method eliminates the need for large training datasets, making advanced reconstruction more accessible.
• This approach provides a flexible alternative for high-resolution imaging in diverse microscopy applications.