Unrolled Richardson-Lucy deconvolution network with partially connected layers in computational microscopy.

Area of Science:

• Computational imaging
• Microscopy
• Deep learning for scientific imaging

Background:

• Computational microscopy utilizes optical encoders with non-localized point spread functions (PSFs) for advanced imaging.
• Multi-cluster PSFs are key in light-field and mask-based integrated microscopes for single-shot 3D imaging.
• Image reconstruction relies on computational decoders, with traditional methods being slow and deep learning lacking interpretability.

Purpose of the Study:

• To develop a computationally efficient and physically interpretable image reconstruction method for microscopy systems with complex PSFs.
• To integrate model-based insights into a data-driven deep learning architecture.
• To address the limitations of traditional iterative methods and deep neural networks in computational microscopy.

Main Methods:

• Proposed a partially connected unrolled Richardson-Lucy network (PC-RLN) architecture.
• Designed each network stage to mimic Richardson-Lucy (RL) deconvolution iterations.
• Incorporated learnable, partially connected layers to model forward imaging and back-projection for non-localized, multi-cluster, and spatially variant PSFs.

Main Results:

• The PC-RLN enables efficient and interpretable image reconstruction.
• Achieved substantially reduced computational cost compared to standard RL deconvolution.
• Demonstrated effectiveness in Fourier light-field microscopy and mask-based integrated microscopy with microlens arrays.

Conclusions:

• The PC-RLN offers a powerful new tool for image reconstruction in advanced microscopy.
• This physics-informed neural network approach balances speed, interpretability, and accuracy.
• The method is suitable for various computational microscopy techniques employing complex PSFs.