Lightweight Richardson-Lucy network for accelerated fluorescence microscope deconvolution.

Area of Science:

• Microscopy and Image Processing
• Computational Imaging
• Biophysics

Background:

• Deconvolution enhances fluorescence microscopy contrast and resolution.
• Deep unfolding deconvolution networks offer improved reconstruction but face computational complexity and data demands.

Purpose of the Study:

• To introduce a lightweight neural network, the accelerated Richardson-Lucy network (ARLN), for efficient fluorescence microscopy deconvolution.
• To improve reconstruction quality and speed while maintaining model interpretability.

Main Methods:

• Developed ARLN by integrating momentum and the Richardson-Lucy model into a lightweight neural network.
• Introduced an additive accelerated vector via minimal trainable convolution layers into the Richardson-Lucy iterative scheme.
• Validated ARLN using both simulated and real fluorescence microscopy data.

Main Results:

• ARLN achieves comparable or superior reconstruction performance, particularly in small-sample scenarios.
• Reduced iterations by 6-10× compared to the accelerated Richardson-Lucy algorithm.
• Achieved >70% reduction in computational cost and parameters versus other lightweight networks.
• Reconstruction speed was 4-12× faster than representative model-driven networks.

Conclusions:

• ARLN offers an interpretable and efficient solution for fluorescence microscopy deconvolution.
• Its lightweight design and accelerated deconvolution capability are valuable for resource-constrained and fast imaging applications.
• ARLN demonstrates significant potential for practical applications in modern fluorescence microscopy.