Dynamical machine learning volumetric reconstruction of objects' interiors from limited angular views.

Area of Science:

• Computational imaging
• Applied physics
• Machine learning

Background:

• Limited-angle tomography is an ill-posed problem requiring regularization for accurate 3D interior volume reconstruction.
• Static neural networks have been used to learn priors for structured objects, but a dynamic approach offers new possibilities.

Purpose of the Study:

• To develop a novel dynamic learning method for limited-angle tomography.
• To adapt nonlinear system identification techniques for improved volumetric reconstruction.
• To validate the dynamic approach for both weak and strong scattering scenarios.

Main Methods:

• Viewed limited-angle tomography as a dynamical system with image acquisition angles as discrete time steps.
• Devised a Recurrent Neural Network (RNN) architecture using a Separable-Convolution Gated Recurrent Unit (SC-GRU).
• Compared the dynamic method against existing approaches using quantitative metrics.

Main Results:

• The dynamic learning method demonstrates suitability for generic interior-volumetric reconstruction under limited-angle conditions.
• Accurate reconstruction of volume interiors was achieved in both weak scattering (Radon transform applicable) and strong scattering (nonlinear) regimes.
• The proposed SC-GRU based RNN effectively regularizes reconstructions.

Conclusions:

• A dynamic learning approach, framed as nonlinear system identification, offers a powerful alternative for limited-angle tomography.
• The SC-GRU based RNN is effective for reconstructing 3D volumes with complex scattering properties.
• This method advances the field of computational imaging for applications in medicine, biology, and industry.