A primal-dual data-driven method for computational optical imaging with a photonic lantern

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning in Imaging

Background:

• High-resolution in vivo imaging is crucial for biological processes.
• Computational Optical Imaging using a Lantern (COIL) offers a single-pixel imaging solution.
• Existing SARA-COIL methods rely on sparsity priors for image reconstruction.

Purpose of the Study:

• To develop a data-driven approach for COIL microendoscopy.
• To improve image reconstruction quality in COIL by integrating learned denoisers.
• To validate the performance of the new method on simulated and real data.

Main Methods:

• Developed a plug-and-play (PnP) algorithm by replacing the sparsity prior with a learned denoiser.
• Utilized a proximal primal-dual algorithm to solve the inverse problem based on Morozov formulation.
• Trained a neural network using learning theory to ensure desirable Lipschitz properties.

Main Results:

• The PnP algorithm demonstrated improved image reconstruction quality compared to the variational SARA-COIL method.
• Convergence of the primal-dual PnP algorithm to a solution of a monotone inclusion problem was shown.
• Successful application on both simulated and real microendoscopy data.

Conclusions:

• The data-driven PnP approach offers a significant advancement in COIL microendoscopy.
• Learned denoisers provide a powerful alternative to traditional sparsity priors for image reconstruction.
• This method enhances the potential of optical fibers for in vivo biological imaging.