Limited angle tomography for transmission X-ray microscopy using deep learning.

Area of Science:

• Nanoscale imaging
• Materials science
• Biomedical imaging

Background:

• Transmission X-ray microscopy (TXM) systems often face sample rotation restrictions, leading to limited angular data.
• Reconstruction from limited angle data in TXM results in artifacts due to missing information.

Purpose of the Study:

• To apply deep learning for the first time to limited angle reconstruction in TXMs.
• To investigate training deep neural networks using synthetic data for artifact reduction.

Main Methods:

• U-Net, a deep neural network, was trained on synthetic ellipsoid and multi-category data.
• The method was evaluated on synthetic data and real chlorella data using 100° limited angle tomography.
• Penalized weighted least-square denoising was applied to measured projections.

Main Results:

• U-Net significantly reduced root-mean-square error (RMSE) and improved the structural similarity (SSIM) index for synthetic data.
• Further improvements in RMSE and SSIM were achieved with penalized weighted least-square denoising.
• Remarkable enhancement in 3D visualization of subcellular structures in real chlorella data was observed.

Conclusions:

• Deep learning, trained on synthetic data, is a viable method for limited angle reconstruction in TXM.
• The proposed U-Net approach significantly reduces artifacts and enhances image quality.
• This technique holds significant value for nanoscale imaging in biology, nanoscience, and materials science.