Zero-shot learning enables instant denoising and super-resolution in optical fluorescence microscopy

Area of Science:

• Microscopy and Imaging Technologies
• Computational Biology and Bioinformatics
• Biophysics

Background:

• Super-resolution microscopy significantly enhances optical imaging capabilities.
• Supervised deep learning models offer high performance but require extensive training data, which is challenging for dynamic biological samples.
• Existing methods struggle with the practical limitations of acquiring high-quality training data for live-cell imaging.

Purpose of the Study:

• To develop an unsupervised deep learning method for super-resolution microscopy that eliminates the need for ground truth data.
• To achieve significant resolution enhancement with reduced fluorescence intensity and without additional data acquisition.
• To demonstrate the broad applicability of the developed method across various microscopy techniques and biological samples.

Main Methods:

• Development of zero-shot deconvolution networks (ZS-DeconvNet), an unsupervised deep learning approach.
• Application of ZS-DeconvNet to enhance resolution beyond the diffraction limit.
• Validation across diverse imaging modalities: TIRF, 3D wide-field, confocal, two-photon, lattice light-sheet, and multimodal SIM.

Main Results:

• ZS-DeconvNet achieved over 1.5-fold resolution enhancement beyond the diffraction limit.
• The method requires 10-fold lower fluorescence intensity compared to ordinary super-resolution conditions.
• Enabled multi-color, long-term, super-resolution 2D/3D imaging of subcellular processes in mouse and C. elegans embryos and mitotic cells.

Conclusions:

• ZS-DeconvNet offers a powerful, unsupervised solution for super-resolution microscopy.
• The method overcomes the limitations of data-hungry supervised approaches for live-cell imaging.
• ZS-DeconvNet provides versatile and efficient super-resolution imaging for diverse biological applications.