ReCSAI: recursive compressed sensing artificial intelligence for confocal lifetime localization microscopy.

Area of Science:

• Super-resolution microscopy
• Biophysics
• Computational imaging

Background:

• Localization microscopy reconstructs nanoscale structures using fluorescent emitter coordinates.
• Traditional methods struggle with irregular point spread functions (PSFs) in techniques like confocal lifetime imaging.
• Compressed sensing and deep learning offer alternatives but have limitations in computational cost and generalization.

Purpose of the Study:

• To develop an efficient method for fitting irregular PSFs in confocal lifetime localization microscopy.
• To combine the strengths of deep learning and compressed sensing for improved reconstruction.
• To enhance the acquisition speed and throughput of super-resolution microscopy.

Main Methods:

• Introduction of ReCSAI, a compressed sensing neural network tailored for confocal dSTORM.
• Development of a simulation tool for generating training data.
• Comparison of various artificial network architectures, focusing on a U-Net with a recursive structure.
• Integration of a trainable wavelet denoising layer.

Main Results:

• The U-Net architecture inspired by iterative compressed sensing demonstrated superior performance on simulated and real confocal lifetime scanning data.
• The addition of a denoising layer further enhanced reconstruction quality.
• ReCSAI achieved reconstruction accuracy comparable to traditional frame binning methods.

Conclusions:

• The deep learning approach enables accurate confocal dSTORM reconstruction without multiple frames.
• The study provides insights into reconstructing sparse, noisy data using combined compressed sensing and deep learning.
• Open-source code and trained networks are provided for reproducibility.