4D-ONIX for reconstructing 3D movies from sparse X-ray projections via deep learning.

Area of Science:

• Physics
• Materials Science
• Computer Science

Background:

• X-ray free-electron lasers and storage rings offer advanced spatiotemporal resolution for studying dynamic processes.
• X-ray multi-projection imaging provides 3D information from single pulses, faster than traditional methods, but faces reconstruction challenges with sparse data.

Purpose of the Study:

• To develop a novel deep learning approach, 4D-ONIX, for reconstructing 4D (3D + time) information from highly sparse X-ray projection data.
• To enable high-fidelity imaging of fast dynamic processes at speeds three orders of magnitude greater than existing techniques.

Main Methods:

• Developed 4D-ONIX, integrating a computational physical model of X-ray matter interaction with advanced deep learning techniques.
• Trained and validated the model using simulations of water droplet collisions and experimental data from additive manufacturing.

Main Results:

• Demonstrated high-quality 4D reconstruction from extremely limited projections (2-3 per timestamp).
• Achieved reconstructions generalized across different experimental conditions.
• Showcased the capability to analyze dynamics at speeds three orders of magnitude faster than conventional tomography.

Conclusions:

• 4D-ONIX is an effective deep learning tool for reconstructing 4D X-ray imaging data from sparse projections.
• The method significantly advances the analysis of fast dynamic processes, overcoming limitations of current reconstruction algorithms.