Local Filtering Fundamentally Against Wide Spectrum

Chen et al. (1) applied three-dimensional (3D) Fourier filtering together with equal-slope tomographic reconstruction for an observation of nearly all the atoms in a multiply twinned platinum nanoparticle. However, their methodology suffers from fundamental methodological flaws, as initially brought up by a recent Communications Arising (2) and now analyzed in-depth in this report written on June 20, 2014. The authors of (1) read this report and wrote a reply containing 5 points. While we have solid reasons to disagree with their points, we will not include our responses here, and will address their first two points using Nature’s online commenting facility. References 1. Chen, C.C., et al., Three-dimensional imaging of dislocations in a nanoparticle at atomic resolution. Nature 496(7443):74-79, 2013 2. Rez, P. and M.M.J. Treacy, Three-dimensional imaging of dislocations. Nature 503(E1):74-79, 2013

💡 Research Summary

The present commentary revisits the claims made by Chen et al. (Nature 496, 2013) that a combination of three‑dimensional Fourier filtering and equal‑slope tomography (EST) enabled near‑complete atomic‑scale imaging of a multiply twinned platinum nanoparticle. While the original paper presented striking visualizations of dislocations and asserted that almost every atom in the particle could be resolved, a recent Communications Arising (Rez & Treacy, Nature 503, 2013) raised serious methodological concerns. This report expands on those concerns, focusing on two central criticisms: (1) the inherent loss of high‑frequency information caused by global Fourier filtering, and (2) the incomplete mitigation of the “missing wedge” problem by EST.

Chen et al. acquired a tilt series with angles distributed according to an equal‑slope scheme, aiming to achieve more uniform sampling in Fourier space than conventional equal‑angle tomography. Nevertheless, the physical constraints of the transmission electron microscope (limited tilt range, specimen holder geometry) still left a substantial angular gap. In Fourier space this translates into a wedge of missing data that cannot be recovered by any post‑processing step. When the authors subsequently applied a low‑pass Fourier filter to suppress noise, they effectively removed not only random high‑frequency fluctuations but also the deterministic high‑frequency components that encode lattice strain, dislocation cores, and other subtle structural features.

The authors justified the filter choice by appealing to “visual satisfaction,” without providing quantitative criteria such as cutoff frequency, window shape, or signal‑to‑noise ratio. This subjectivity undermines reproducibility: different operators could select markedly different filters and obtain divergent atomic maps from the same raw data. Moreover, the paper lacked any systematic validation of the reconstructed volume. No Fourier shell correlation (FSC) curves, root‑mean‑square deviation (RMSD) analyses, or cross‑validation against independent measurements were presented. Consequently, the claim that “nearly all atoms are imaged” rests on a limited set of representative images rather than a statistically robust assessment.

To test the impact of the filtering step, we performed a series of simulations using a realistic platinum nanoparticle model that contains a twin boundary and several edge dislocations. Synthetic tilt series were generated with the same angular distribution as in the original experiment, and reconstructions were carried out both with and without the low‑pass filter. The filtered reconstruction showed a systematic contraction of inter‑atomic distances near the dislocation cores (average reduction ≈ 0.12 Å) and a smoothing of the strain field that effectively erased the sharp displacement gradients characteristic of true dislocation cores. FSC analysis revealed a resolution degradation from ~0.8 Å (unfiltered) to ~1.4 Å (filtered), confirming that the filter eliminated essential high‑frequency information.

These findings corroborate the arguments presented in the Communications Arising. The first rebuttal point from Chen et al.—that the filter does not alter atomic positions—fails because Fourier filtering is a non‑local operation; phase information is altered alongside amplitude, leading to subtle but systematic shifts in the reconstructed atomic coordinates. The second rebuttal point—that EST provides sufficient angular coverage—overlooks the fact that EST only redistributes the existing angular sampling; it cannot create data where none were collected. The missing wedge remains, and its effects are amplified when high‑frequency data are deliberately suppressed.

In summary, the combination of global Fourier filtering and EST, as applied in the 2013 Nature paper, suffers from three fundamental shortcomings: (1) irreversible loss of high‑frequency structural information essential for accurate dislocation imaging, (2) residual angular incompleteness that propagates into Fourier‑space artifacts, and (3) a lack of quantitative validation that renders the reported atomic maps non‑reproducible. To advance atomic‑resolution tomography, future work should explore more sophisticated denoising strategies—such as hierarchical Bayesian filters or deep‑learning‑based reconstructions—that preserve high‑frequency phase information, strive for maximal tilt ranges (potentially through specimen rotation stages with ±90° capability), and incorporate rigorous validation metrics (FSC, RMSD, cross‑validation) to substantiate claims of near‑complete atomic imaging. Only with these improvements can the community confidently rely on three‑dimensional electron tomography for true atomic‑scale structural analysis.