TL;DR
This paper introduces a "quasi-realtime" X-ray microtomography system at the Advanced Photon Source that significantly speeds up the process of acquiring and reconstructing 3D images. By pipelining data acquisition, high-speed transfer, and parallel computation, they achieve near-instantaneous feedback on tomographic results, enabling interactive experimentation.
What Problem Does This Paper Address?
Traditional X-ray microtomography, while powerful for generating high-resolution 3D images, suffers from long acquisition and reconstruction times. This delay hinders interactive experimentation, making it difficult for researchers to quickly assess their sample, adjust parameters, or even confirm that the experiment is proceeding as expected. For user-shared facilities like synchrotron radiation sources, where beamtime is precious, this inefficiency is particularly problematic. The authors aim to bridge this gap by creating a system that provides rapid feedback, allowing for more efficient and dynamic experimental workflows.
Methods
The core of their solution is a highly integrated, pipelined system.
- Data Acquisition: They use a straight-ray projection imaging scheme at the 2-BM beamline, employing a CdWO4 scintillator to convert X-rays to visible light, which is then imaged onto a CCD detector. This system can acquire projections at resolutions of 1-10 µm.
- High-Speed Data Transfer: Immediately after each projection is acquired and saved locally, it's rapidly transferred via a 100 Mb/s network to a remote SGI Origin 2000 parallel computer.
- Parallel Preprocessing and Reconstruction: The parallel computer, utilizing the Globus metacomputing toolkit, performs preprocessing (like white field removal, filtering, and alignment) and reconstruction concurrently with ongoing data acquisition.
- Visualization: The reconstructed data is then sent to a visualization computer (Viz Engine) for volume rendering. This allows operators at the beamline to view rendered images within minutes of data acquisition completion.
- Data Format: They use the Hierarchical Data Format (HDF) with NeXus conventions for organizing experimental data, ensuring platform independence and self-description.
Datasets
The paper describes experiments using two types of samples:
- A micromachined gear (approximately 100 µm diameter), acquired using 15 keV X-rays in phase contrast.
- A carpenter ant's head, acquired using 9 keV X-rays in absorption-contrast mode.
These are not presented as benchmark datasets but rather as examples demonstrating the system's capabilities.
Key Insights and Ideas
- Pipelining is Key: The most significant insight is that by tightly integrating acquisition, transfer, and parallel processing, the overall time to get a reconstructed 3D image can be drastically reduced, approaching the acquisition time itself.
- "Quasi-Realtime" is Sufficient: For many experimental scenarios, achieving results within minutes, rather than seconds, is sufficient to enable interactive adjustments and feedback, revolutionizing experimental efficiency.
- Leveraging Supercomputing: The paper highlights the critical role of parallel computing resources in handling the massive data volumes generated by modern X-ray sources for rapid reconstruction.
- Data Format Matters for Integration: The choice of HDF and NeXus conventions facilitates the seamless transfer and interpretation of data across different computational stages.
Key Results
- A complete tomographic dataset can be collected in approximately 15 minutes.
- Reconstruction results are available for viewing minutes after data acquisition is complete.
- For test runs with 180 projections (512x512 pixels), reconstruction on 20 out of 128 processors took less than 10 minutes.
- When acquiring data in cycles (e.g., 12-24 projections), reconstruction results can be viewed at the experiment station within 2-3 minutes.
- The system currently achieves resolutions of 1-10 µm, with plans to reach below 100 nm.
Limitations and Future Work
- Resolution Limits: Current resolution is around 1-10 µm, with plans to improve this to sub-100 nm by using focused beams and advanced optics.
- Acquisition Time: While significantly reduced, acquisition time is still a bottleneck. Future work aims to reduce it to seconds by improving detectors and using more efficient monochromators.
- Computation Time: Computation time is also being targeted for reduction to under 1 minute through software optimization and parallelization.
- Subsampling for Visualization: The Viz Engine has a data volume size limitation (256x256x256 pixels), requiring subsampling of larger reconstructions for real-time visualization.
- User Interface: A user-friendly graphical user interface (GUI) is under development to allow general users to control the entire pipeline interactively from the beamline.
- Detector Improvements: A scintillator better matched for 5-20 keV X-ray absorption and a high numerical aperture objective are needed to improve efficiency and resolution.
Summary updated