Diffraction tomography with a deep image prior.

Area of Science:

• Optics and Photonics
• Biomedical Imaging
• Computational Imaging

Background:

• 3D refractive index (RI) mapping is crucial for understanding biological sample structure.
• Standard tomographic reconstruction methods struggle with thick samples and the missing cone problem, limiting resolution and quality.
• Deep learning approaches offer potential for improved image reconstruction but often require extensive pre-training.

Purpose of the Study:

• To introduce Deep Prior Diffraction Tomography (DP-DT), a novel technique for high-resolution 3D RI reconstruction of thick biological samples.
• To address the limitations of existing 3D reconstruction algorithms, particularly the missing cone problem.
• To develop a generalizable imaging technique applicable to diverse samples without dataset bias.

Main Methods:

• DP-DT utilizes a phase retrieval algorithm extended by a deep image prior (DIP).
• A deep generative 3D convolutional neural network (CNN) reparameterizes the 3D sample reconstruction.
• The technique processes multi-angle, low-resolution images from angularly varying illumination.

Main Results:

• DP-DT effectively reconstructs high-resolution 3D RI maps from low-resolution, multi-angle data.
• The method successfully addresses the missing cone problem, enhancing image quality and resolution.
• DP-DT demonstrated superior performance compared to standard regularization techniques in both simulated and experimental data.
• The technique showed generality across different scattering models (first Born and multi-slice).

Conclusions:

• DP-DT offers a powerful, generalizable solution for high-resolution 3D RI mapping of thick biological samples.
• The deep image prior approach overcomes limitations of traditional methods and avoids dataset bias.
• DP-DT holds significant potential for advancing various 3D imaging modalities beyond its current application.