Learned SPARCOM: unfolded deep super-resolution microscopy.

Area of Science:

• Optics and Photonics
• Biophysics
• Computational Biology

Background:

• Super-resolution microscopy techniques like single-molecule localization microscopy (SMLM) offer high spatial resolution.
• Traditional SMLM methods often require long acquisition times, limiting their use in dynamic biological processes.
• Achieving high temporal resolution typically compromises spatial precision due to low emitter density.

Purpose of the Study:

• To develop a novel deep learning approach for super-resolution imaging that overcomes the trade-off between temporal and spatial resolution.
• To integrate domain knowledge of optical imaging into a compact neural network architecture.
• To enhance the efficiency and applicability of SMLM for live-cell imaging.

Main Methods:

• Combining SPARCOM (Sparse Reconstruction for Optical Microscopy), a classical super-resolution method, with model-based deep learning.
• Utilizing an algorithm unfolding approach to design a compact neural network (learned SPARCOM or LSPARCOM).
• Training and testing the LSPARCOM network on various datasets, evaluating its performance without prior knowledge of the optical system.

Main Results:

• LSPARCOM successfully achieved super-resolution imaging from a reduced number of high emitter density frames.
• The method demonstrated robustness across different test sets and did not require prior knowledge of the specific optical system.
• The developed network effectively incorporated domain knowledge, leading to improved imaging performance.

Conclusions:

• LSPARCOM offers a promising solution for interpretable and efficient live-cell imaging.
• The integration of classical methods with deep learning provides a powerful framework for advancing microscopy.
• This approach has broad potential applications in single-molecule localization microscopy for studying biological structures.