Real-time undersampling optimization during electron tomography of beam-sensitive samples using golden ratio scanning and RECAST3D

T. Craig, Ajinkya Kadu, Kees Joost Batenburg, Sara BalsBIO 2024 · DOI

TL;DR

This paper introduces a real-time workflow for electron tomography that optimizes the number of projections on the fly. By combining "golden ratio scanning" with quasi-3D reconstruction, researchers can monitor image quality during acquisition and stop the moment they reach the "sweet spot," protecting delicate samples from unnecessary beam damage.

What Problem Does This Paper Address?

Electron tomography (ET) is the gold standard for characterizing the 3D structure of nanomaterials, essentially acting as a CT scan at the atomic or molecular scale. However, ET has a "Goldilocks" problem:

  1. Too many projections: The sample is exposed to a high dose of electrons, which can melt, shrink, or destroy "beam-sensitive" materials like catalysts or metal-organic frameworks (MOFs).
  2. Too few projections: The resulting 3D reconstruction is riddled with artifacts and noise, making it scientifically useless.

Traditionally, finding the optimal number of projections (undersampling optimization) requires multiple trial-and-error acquisitions and lengthy post-processing. This is not only time-consuming but often impossible for rare or highly sensitive samples that degrade after a single scan.

Methods

The authors propose a novel workflow that moves optimization from the "post-processing" phase into "real-time." It relies on two key technologies:

  • Golden Ratio Scanning (GRS): Unlike traditional linear scanning (e.g., -60° to +60° in 2° increments), GRS chooses angles based on the golden ratio. This ensures that no matter when you stop the scan, the angles you've already collected are distributed as uniformly as possible.
  • RECAST3D: This software performs "quasi-3D" reconstruction in real-time. Instead of trying to reconstruct the entire 3D volume (which is computationally expensive), it reconstructs three 2D "orthoslices" (cross-sections) that the researcher can watch as the data comes in.

The Optimization Loop: As the microscope collects data, the system calculates two metrics:

  1. Signal-to-Noise Ratio (SNR): To ensure the image is clear.
  2. Convergence: Measuring how much the reconstruction changes as new projections are added. Once the image stops significantly changing, adding more projections only adds beam damage without adding new information. At this point, the acquisition is terminated.

Datasets

The researchers validated their approach using three distinct types of data:

  • Simulated Structures: A digital model designed to undergo simulated beam damage to test the algorithm's ability to "outrun" the destruction.
  • Au (Gold) Nanoparticle: A stable experimental standard used to compare the new method against traditional tomographic procedures.
  • Beam-Sensitive MOFs: The real-world test cases, specifically Au@NU-1000 and Au/Pd@ZIF-8. These materials are notoriously difficult to image because their porous structures collapse under prolonged electron exposure.

Key Insights and Ideas

  • Dynamic Termination: The biggest "aha moment" is the shift from fixed-length scans to metric-driven scans. By letting the data quality dictate the end of the experiment, researchers avoid the "over-exposure" that typically ruins sensitive samples.
  • The Power of GRS: Because GRS fills the angular space efficiently at every step, the acquisition is "interruptible." You don't need to know the total number of projections at the start.
  • Efficiency over Perfection: By focusing on orthoslices (RECAST3D) rather than a full 3D volume during the scan, the team achieved the speed necessary for real-time feedback without needing a supercomputer.

Key Results

  • Higher Fidelity: In simulated and experimental tests, the real-time optimized reconstructions showed significantly higher fidelity to the original structure compared to standard incremental acquisition.
  • Time Savings: The workflow achieves results comparable to "conventionally optimized" tomography (which takes several acquisitions and hours of post-processing) in a single acquisition.
  • MOF Success: The technique successfully optimized imaging conditions for Au@NU-1000 and Au/Pd@ZIF-8, providing clear 3D data of these delicate materials before beam damage could obscure the results.

Limitations and Future Work

  • Quasi-3D vs. Full 3D: The real-time feedback is currently based on orthoslices. While these are excellent proxies for quality, some fine 3D details might still require full volume reconstruction to be fully appreciated.
  • Hardware Integration: The paper focuses on the workflow and software; however, implementing this requires tight integration between the microscope's control system and the reconstruction software, which may be a hurdle for some labs.
  • Future Directions: The authors suggest this technique is a stepping stone toward fully automated, "smart" microscopy where the instrument itself decides how to best preserve a sample while maximizing data yield.

Summary updated

BibTeX
@inproceedings{craig2024real,
  author = {T. Craig and Ajinkya Kadu and Kees Joost Batenburg and Sara Bals},
  title = {{Real-time undersampling optimization during electron tomography of beam-sensitive samples using golden ratio scanning and RECAST3D}},
  booktitle = {BIO},
  year = {2024},
  doi = {10.1051/bioconf/202412902011},
}
Figure