ReCSAI

ReCSAI reconstructs fluorescent emitter coordinates from diffraction-limited spots in confocal lifetime localization microscopy by combining compressed sensing and deep learning to handle irregular and non-linear point spread functions (PSFs).

Key Features:

• Emitter reconstruction: Reconstructs fluorescent emitter coordinates from diffraction-limited spots in confocal lifetime scanning data.
• PSF robustness: Handles irregular and non-linear point spread functions (PSFs) encountered in confocal lifetime imaging with laser scanning across samples.
• Compressed sensing + deep learning: Integrates compressed sensing techniques with deep learning to exploit sparsity in emitter distributions and model arbitrary PSFs.
• Recursive U-Net architecture: Uses a U-Net with a recursive structure inspired by iterative compressed sensing.
• Trainable wavelet denoising layer: Incorporates a trainable wavelet denoising layer that acts as an explicit prior to improve reconstruction quality.
• Single-frame accuracy: Achieves reconstruction accuracy comparable to traditional fitting methods that require frame binning, without necessitating multiple frames, enabling higher acquisition throughput.
• Simulation and training resources: Includes a simulation tool implemented in Python and Jupyter notebooks, plus code for network training and inference and trained networks for generating simulated training data.

Scientific Applications:

• Confocal lifetime localization microscopy: Reconstruction of emitter coordinates from diffraction-limited measurements in confocal lifetime imaging experiments.
• Sparse emitter recovery: Localization of sparse emitter sets from noisy data using compressed sensing priors.
• Non-linear PSF fitting: Fitting and localization in the presence of irregular, non-symmetric PSFs without strong parametric PSF assumptions.
• High-throughput single-frame localization: Enabling accurate localization without frame binning to increase acquisition speed and throughput in super-resolution experiments.
• Simulation-driven training: Generation of realistic simulated datasets for network training and validation in super-resolution microscopy workflows.

Methodology:

Combines compressed sensing techniques with deep learning via a recursive U-Net architecture inspired by iterative compressed sensing, incorporates a trainable wavelet denoising layer, and was evaluated on realistic simulated datasets (including noise) and real experimental confocal lifetime scanning data; a Python/Jupyter simulation tool and code for network training and inference are provided.