Tomographic reconstruction with a generative adversarial network.

Area of Science:

• Computational imaging
• Medical physics
• Artificial intelligence in science

Background:

• Tomographic reconstruction is crucial for imaging but often faces challenges with limited data.
• Traditional methods can struggle with ill-posed inverse problems, particularly in missing-wedge tomography.

Purpose of the Study:

• To develop a novel deep learning algorithm, GANrec, for direct tomographic reconstruction.
• To enhance reconstruction accuracy and quality, especially for datasets with limited angular information.

Main Methods:

• Utilized a generative adversarial network (GAN) to directly solve the inverse Radon transform.
• Implemented a self-training procedure based on a physics model, fitting input sinograms to model sinograms.
• Applied the algorithm to missing-wedge tomography and X-ray ptychographic tomography (PXCT) data.

Main Results:

• Achieved significant improvements in reconstruction accuracy for limited-angle and missing-wedge tomography.
• Successfully reconstructed the 3D pore structure of a zeolite particle from sparse PXCT data (51 projections over 70°).
• Demonstrated the algorithm's effectiveness without requiring additional training steps for independent sinograms.

Conclusions:

• GANrec offers a robust and accurate method for tomographic reconstruction, particularly in data-scarce scenarios.
• The self-training, physics-based approach provides a universal reconstruction concept for ill-posed inverse problems with well-defined forward models.