Physics-guided self-supervised learning for high-fidelity reconstruction in live-cell imaging.

Area of Science:

• Microscopy and Imaging
• Biophysics
• Computational Biology

Background:

• Long-term live-cell imaging necessitates low excitation intensity to minimize phototoxicity.
• Low excitation intensity results in insufficient signal-to-noise ratios (SNR), complicating image reconstruction.
• Existing methods often fail to balance noise reduction and fine structure preservation, leading to artifacts.

Purpose of the Study:

• To develop a robust reconstruction framework for super-resolution deconvolution imaging under low-SNR conditions.
• To enable high-fidelity image reconstruction without requiring ground truth data.
• To improve the visualization of dynamic subcellular processes in live cells.

Main Methods:

• Introduced HiFi-DeconvFormer, a physics-guided self-supervised reconstruction framework.
• Employed a spatial-redundancy-based training strategy combined with physical imaging models.
• Utilized windowed Transformers and multi-modal regularization for feature extraction.

Main Results:

• Achieved robust self-supervised learning without ground truth data.
• Effectively captured long-range biological continuity and high-frequency edge features.
• Demonstrated superior performance over state-of-the-art methods in low-SNR conditions.
• Enabled artifact-free reconstruction of dynamic subcellular processes.

Conclusions:

• HiFi-DeconvFormer provides high-fidelity reconstructions for challenging live-cell imaging scenarios.
• The framework successfully recovers weak signals and preserves fine subcellular structures.
• This method advances the capability for detailed observation of dynamic biological events in live cells.