ML-SIM: universal reconstruction of structured illumination microscopy images using transfer learning

Area of Science:

• Optical microscopy
• Super-resolution imaging
• Computational imaging

Background:

• Structured illumination microscopy (SIM) offers high-resolution live-cell imaging.
• Traditional SIM reconstruction is slow, parameter-dependent, and prone to artifacts.
• Existing methods require specific tuning for different hardware and sample types.

Purpose of the Study:

• To develop a versatile, parameter-free reconstruction method for SIM.
• To improve the speed, robustness, and generalizability of SIM image reconstruction.
• To enable efficient super-resolution imaging under challenging experimental conditions.

Main Methods:

• Introduced ML-SIM, an end-to-end deep residual neural network.
• Utilized transfer learning with simulated data for training.
• Trained on challenging conditions to ensure robustness to noise and illumination irregularities.
• Applied the model to diverse experimental SIM datasets from multiple microscopes.

Main Results:

• ML-SIM achieved high-quality reconstructions across various sample types and microscopes.
• Demonstrated superior generality and robustness to noise compared to state-of-the-art methods.
• Reconstruction of a SIM stack completed in under 200 ms on a GPU.
• The method is parameter-free and does not require experimental training data.

Conclusions:

• ML-SIM provides a versatile and efficient alternative for SIM image reconstruction.
• The deep learning approach overcomes limitations of traditional methods, enhancing applicability.
• Fast reconstruction speeds pave the way for real-time super-resolution imaging applications.