Distributed Reconstruction Algorithm for Electron Tomography with Multiple-scattering Samples

David Ren, Michael Whittaker, Colin Ophus, Laura WallerarXiv 2021 · DOI · arXiv

Abstract

Three-dimensional electron tomography is used to understand the structure and properties of samples in chemistry, materials science, geoscience, and biology. With the recent development of high-resolution detectors and algorithms that can account for multiple-scattering events, thicker samples can be examined at finer resolution, resulting in larger reconstruction volumes than previously possible. In this work, we propose a distributed computing framework that reconstructs large volumes by decomposing a projected tilt-series into smaller datasets such that sub-volumes can be simultaneously reconstructed on separate compute nodes using a cluster. We demonstrate our method by reconstructing a multiple-scattering layered clay (montmorillonite) sample at high resolution from a large field-of-view tilt-series phase contrast transmission electron microscopy dataset.

Introduction: The Challenge of "Thick" Samples

In traditional Electron Tomography (ET), the "weak phase approximation" assumes electrons pass through a sample linearly, like light through thin glass. However, as samples increase in thickness or atomic weight, electrons undergo multiple scattering, bouncing off atoms several times. Accounting for these nonlinear physics requires complex "multi-slice" models that are computationally intensive. Modern 8k detectors generate datasets so massive they exceed the memory capacity of single high-end workstations, and existing distributed methods often fail because they assume linear physics. This research introduces a "divide and conquer" framework that preserves nonlinear scattering physics while enabling massive parallelization.

The Distributed Framework

The core innovation of this approach is a mathematical decomposition that allows a sub-volume within a larger sample to be reconstructed using only a specific subset of the original tilt-series.

Overview of the distributed reconstruction workflow using the 'Shift and Crop' method.

The "Shift and Crop" Method

By shifting each raw image in the tilt-series by a calculated amount—$\Delta x_\theta = x_0 \cos(\theta) + z_0 \sin(\theta)$—and then cropping it, the researchers create a "mini tilt-series." This subset behaves as if the sub-volume were the only object being imaged. Because these datasets are self-contained, different compute nodes in a cluster can process them simultaneously without needing to communicate during the reconstruction, eliminating a common bottleneck in distributed systems.

Optimization and Performance Trade-offs

Implementing this framework requires balancing reconstruction accuracy against computational cost. A key finding was the 30% Overlap Rule: to avoid edge artifacts caused by scattering from outside the sub-volume, adjacent regions must overlap significantly. This is derived from the geometry of an inscribed cuboid within the cylindrical support of a tomographic reconstruction.

Analysis of reconstruction quality and computational efficiency versus sub-volume size.

As shown in the performance analysis, smaller sub-volumes allow for higher parallelization (more nodes) and faster individual node processing, but they introduce more artifacts and higher total computational overhead due to the required overlap. The authors identified a sub-volume size of $150^3$ voxels as an efficient "sweet spot." Interestingly, the study also considered environmental impact, noting that while smaller sub-volumes are faster in terms of wall-clock time, they consume more total energy (and thus have a higher carbon footprint) due to the redundant processing of overlapping regions.

Key Results and Impact

The algorithm was validated using a tilt-series of lithium-montmorillonite clay in a frozen electrolyte solution, captured on a Titan Krios TEM.

  • Record-Breaking Scale: The team produced a volume of $3952 \times 5500 \times 952$ voxels, which was the largest cryo-ET reconstruction at the time of publication.
  • High Resolution: They achieved an isotropic resolution of 1.82 ų/voxel, resolving individual clay layers separated by only ~1nm—features previously obscured by blurring in lower-resolution models.
  • Algorithmic Flexibility: Because the "shift and crop" preprocessing occurs before reconstruction, this framework is compatible with any reconstruction algorithm, whether linear or nonlinear.

Conclusion and Future Work

While the method successfully scales electron tomography for thick samples, it faces trade-offs. Smaller sub-volumes eventually lose too much "global" scattering information, leading to artifacts. Future work may explore more advanced "content-aware" stitching techniques to further reduce boundary artifacts between sub-volumes. Nonetheless, this distributed framework provides a vital pathway for high-resolution 3D imaging of complex, large-scale materials science and biological samples.

Summary updated

Abstract

Three-dimensional electron tomography is used to understand the structure and properties of samples in chemistry, materials science, geoscience, and biology. With the recent development of high-resolution detectors and algorithms that can account for multiple-scattering events, thicker samples can be examined at finer resolution, resulting in larger reconstruction volumes than previously possible. In this work, we propose a distributed computing framework that reconstructs large volumes by decomposing a projected tilt-series into smaller datasets such that sub-volumes can be simultaneously reconstructed on separate compute nodes using a cluster. We demonstrate our method by reconstructing a multiple-scattering layered clay (montmorillonite) sample at high resolution from a large field-of-view tilt-series phase contrast transmission electron microscopty dataset.
BibTeX
@misc{ren2021distributed,
  author = {David Ren and Michael Whittaker and Colin Ophus and Laura Waller},
  title = {{Distributed Reconstruction Algorithm for Electron Tomography with Multiple-scattering Samples}},
  year = {2021},
  eprint = {2110.07857},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.07857},
}

Loading related content...

Figure