Abstract
Tomographic X-ray microscopy beamlines at synchrotron light sources worldwide have pushed the achievable time-resolution for dynamic 3-dimensional structural investigations down to a fraction of a second, allowing the study of quickly evolving systems. The large data rates involved impose heavy demands on computational resources, making it difficult to readily process and interrogate the resulting volumes. The data acquisition is thus performed essentially blindly. Such a sequential process makes it hard to notice problems with the measurement protocol or sample conditions, potentially rendering the acquired data unusable, and it keeps the user from optimizing the experimental parameters of the imaging task at hand. We present an efficient approach to address this issue based on the real-time reconstruction, visualisation and on-the-fly analysis of a small number of arbitrarily oriented slices. This solution, requiring only a single additional computing workstation, has been implemented at the TOMCAT beamline of the Swiss Light Source. The system is able to process multiple sets of slices per second, thus pushing the reconstruction throughput on the same level as the data acquisition. This enables the monitoring of dynamic processes as they occur and represents the next crucial step towards adaptive feedback control of time-resolved in situ tomographic experiments.
Introduction: The Challenge of "Blind" Acquisition
In modern synchrotron facilities, X-ray detectors have become incredibly fast, capable of capturing the data needed for a 3D snapshot in a fraction of a second. However, the computational "back end" hasn't kept pace. Traditionally, reconstructing a high-resolution 3D volume takes minutes or even hours. This creates a "blind" acquisition process where scientists often don't know if their sample is perfectly aligned, if the region of interest is in view, or if the dynamic process they are studying has even started until after the experiment is over. If something goes wrong, the data—and the expensive beamtime used to get it—is wasted. This research addresses the need for immediate, "live" visual feedback to enable adaptive control of experiments.
Methods: A High-Speed Pipeline for Real-Time Feedback
The researchers developed a pipeline integrated at the TOMCAT beamline (Swiss Light Source) that bridges the gap between high-speed acquisition and slow reconstruction. The hardware utilizes the GigaFRoST detector system, which streams raw data at a massive 7.7 GB/s via fiber optics to a backend server.
The core innovation is the shift from full 3D reconstruction to "Quasi-3D." Instead of trying to reconstruct a full $N^3$ volume, which is computationally heavy, the system uses the RECAST3D framework to reconstruct a small number of user-defined, arbitrarily oriented 2D slices. The system uses Filtered Backprojection (FBP); because FBP is linear and the contribution of each voxel is independent, the pipeline performs row-based filtering on CPUs and then uses a single GPU to backproject only the specific pixels needed for the chosen slices. The architecture leverages ZeroMQ for distributed message passing, allowing a single workstation to subscribe to the live data stream, process it, and update the visualization in real-time.
Key Insights: Slices as Proxies for Control
A central insight of this work is that a full 3D volume is not strictly necessary to steer an experiment. Three orthogonal slices (XY, XZ, YZ) that can be tilted and moved in real-time provide sufficient spatial context for the majority of experimental decision-making. Reconstructing a slice is $O(n^2 N_\phi)$ whereas a volume is $O(n^3 N_\phi)$; this reduction in complexity is what makes "real-time" possible on a modest single-GPU workstation.
Furthermore, the pipeline is designed to be "lossy" but robust. If the reconstruction server cannot keep up with the 7.7 GB/s stream, it drops frames. The resulting "missing angle" artifacts are acceptable for a live preview, as the full-quality data is simultaneously saved to disk for rigorous post-experiment analysis.
Experimental Validation and Performance
The pipeline was validated using real-world dynamic experiments, such as tracking water wicking in a bundle of 96 polyethylene terephthalate (PET) fibers and navigating the internal structure of volcanic rock.
Key performance metrics include:
- Latency: Total response time—from user interaction to updated reconstruction—is less than 100ms.
- Throughput: The system successfully processed 400 projections and updated three slices in roughly 728ms, matching sub-second acquisition rates.
- Practical Utility: The system enabled live alignment (correcting rotation axis offsets in seconds), live positioning of microscopic regions of interest, and live observation of fast-moving fronts.
Future Directions: Towards Autonomous Control
While the current system provides a vital "eye" for researchers, future work aims to move beyond human-in-the-loop control. The authors envision autonomous feedback where the system triggers actions—such as starting a pump or stopping the beam—based on automated image analysis of the live slices. Additionally, by scaling from a single workstation to a small cluster of approximately 8 GPU nodes, full-resolution, real-time 3D reconstruction is expected to become feasible.
Summary updated