Zero-Shot Learning for Limited Photon Budget Denoising in Structured Illumination Microscopy

Area of Science:

• Microscopy
• Image Processing
• Computational Biology

Background:

• Structured illumination microscopy (SIM) enables rapid live-cell imaging but suffers from noise and artifacts.
• Low photon efficiency in SIM acquisition exacerbates image degradation, hindering dynamic cellular process investigation.

Purpose of the Study:

• To develop a novel, efficient, and accurate denoising method for low-photon efficiency SIM images.
• To address artifacts and noise interference in SIM image reconstruction.

Main Methods:

• Proposed a zero-shot learning-based SIM image denoising method (ZS-SIM) utilizing neural network training on a single noisy acquisition.
• Integrated traditional Wiener-SIM reconstruction for physical fidelity and employed downsampling/interpolation for resampling.
• Introduced symmetric reconstruction loss and mutual constraint SSIM loss to improve training stability and convergence speed.

Main Results:

• ZS-SIM achieved artifact-free, high-fidelity denoising reconstruction with low model complexity and fast inference.
• Demonstrated effective denoising for low-photon efficiency live-cell imaging and scenarios with limited computational resources.
• Validated ZS-SIM's effectiveness for scanning electron microscopy (SEM) data denoising, improving downstream segmentation tasks.

Conclusions:

• ZS-SIM offers a practical solution for rapid, high-quality SIM image reconstruction in low-light conditions.
• The method shows potential for advancing low-photon efficiency imaging and supporting rapid validation in biomedical research.
• ZS-SIM's applicability to SEM data suggests broader utility in various imaging modalities.