High-resolution limited-angle phase tomography of dense layered objects using deep neural networks

Area of Science:

• Physics
• Computer Science
• Materials Science

Background:

• Phase tomography typically requires multiple steps, involving separate phase retrieval and reconstruction.
• Limited projection angles and low photon flux pose challenges for traditional tomographic reconstruction.

Purpose of the Study:

• To develop a single-step, machine learning-based method for 3D tomographic reconstruction.
• To enable reconstruction from intensity projections directly, bypassing intermediate phase retrieval.
• To demonstrate the method's effectiveness under challenging conditions like limited angles and low photon flux.

Main Methods:

• A physics-informed preprocessor followed by a deep neural network (DNN) was employed.
• The method directly reconstructs 3D data from intensity projections.
• A scaled-up integrated circuit phantom was used for experimental validation in the visible optical domain.

Main Results:

• Successful 3D reconstruction of dense layered objects was achieved with a limited range of projection angles.
• The DNN, trained solely on synthetic data, accurately reconstructed physical samples.
• The method demonstrated robustness even under highly attenuated photon fluxes.

Conclusions:

• The presented single-step method simplifies and enhances tomographic reconstruction.
• Machine learning, particularly DNNs trained on synthetic data, can overcome limitations of traditional methods.
• The approach is broadly applicable across various radiation types and bands for tomographic imaging.