3D Gaussian Splatting for Annular Dark Field Scanning Transmission Electron Microscopy Tomography Reconstruction

Abstract

Analytical Dark Field Scanning Transmission Electron Microscopy (ADF-STEM) tomography reconstructs nanoscale materials in 3D by integrating multi-view tilt-series images, enabling precise analysis of their structural and compositional features. Although integrating more tilt views improves 3D reconstruction, it requires extended electron exposure that risks damaging dose-sensitive materials and introduces drift and misalignment, making it difficult to balance reconstruction fidelity with sample preservation. In practice, sparse-view acquisition is frequently required, yet conventional ADF-STEM methods degrade under limited views, exhibiting artifacts and reduced structural fidelity. To resolve these issues, in this paper, we adapt 3D GS to this domain with three key components. We first model the local scattering strength as a learnable scalar field, denza, to address the mismatch between 3DGS and ADF-STEM imaging physics. Then we introduce a coefficient $γ$ to stabilize scattering across tilt angles, ensuring consistent denza via scattering view normalization. Finally, We incorporate a loss function that includes a 2D Fourier amplitude term to suppress missing wedge artifacts in sparse-view reconstruction. Experiments on 45-view and 15-view tilt series show that DenZa-Gaussian produces high-fidelity reconstructions and 2D projections that align more closely with original tilts, demonstrating superior robustness under sparse-view conditions.

Figures (6)

• Figure 1: Projection comparison under 15-view reconstruction across traditional reconstruction methods and our approach. Brightness is enhanced while contrast is reduced to highlight artifacts. FDK exhibits the most severe artifacts, followed by SIRT and GENFIRE. Our method yields the cleanest projections.
• Figure 2: Pipeline of Denza-Gaussian. DenZa-Gaussian is a two-stage Architecture tailored for sparse-view ADF-STEM tomography. The first stage uses FDK initialization to rapidly generate a rough 3D volume capturing the sample’s overall structural topology. The second stage converts this rough volume into a point cloud for Gaussian initialization. We use a custom 2D rendering pipeline optimized for ADF-STEM’s cone-beam imaging geometry to project the initialized 3D Gaussian cloud into 2D images. Loss is computed between these projections and ground truth images, complemented by regularization constraints on the 3D volume. Iterative optimization delivers a high-quality 3D reconstruction.
• Figure 3: Results of qualitative comparison under 45-view. We performed 3-D reconstruction of PtNi nanomaterial blocks from 45 tilt-series projections. Compared methods are SIRT andersen1984simultaneous, FDK feldkamp1984practical, GENFIRE pryor2017genfire, and 3D GS kerbl20233d.
• Figure 4: Results of qualitative comparison under 15-view. We performed 3-D reconstruction of PtNi nanomaterial blocks from 15 tilt-series projections. Compared methods are SIRT andersen1984simultaneous, FDK feldkamp1984practical, GENFIRE pryor2017genfire, and 3D GS kerbl20233d.
• Figure 5: Comparison of 3D volume reconstruction results. The left panel compares the reconstructed 3D volume surfaces, while the right panel focuses on edge reconstruction. Part (a) presents a comparison between GENFIRE and our method on full 3D volumes, and part (b) shows an ablation study of the 2D Fourier loss under sparse-view conditions.