Abstract text (incl. figure legends and references)

Electron tomography is a valuable tool to obtain the 3D structure of specimen in the nm or even sub-nm range. The general workflow is to acquire a series of projections from the specimen from different angles or orientations, computationally reconstruct the volume and segment the volume data into different components or phases. Over the last decades a variety of reconstruction, segmentation and filtering algorithms have emerged, providing the scientist with a large toolset for his tasks. However, by favoring one algorithm over the other, the resulting tomographic reconstruction may deviate from the true 3D structure.

In this work we present the influence of reconstruction and segmentation algorithms on the final 3D structure of a nanoporous silicon sample with known porosity.

The sample material consist of electrochemically anodized silicon with cylindrical pores of a diameter <10 nm and a solid-fraction of 50% determined by nitrogen sorption isotherm. A tomography needle of 150 nm diameter was prepared by standard FIB liftout technique. A tilt series of 161 projections (‑80° ‑ +80° with 1° spacing) was acquired using a Talos F200X TEM in HAADF-STEM mode.

After image registration of the unfiltered projections, the volume was reconstructed using WBP, SIRT, SART and EM algorithms. A representative volume of the reconstructed volume was chosen for histogram based thresholding segmentation using Otsu, Max Entropy, K-means and Percentile (50%).

The solid-void ratio, homogeneity of density and noise of the segmented volume were analyzed and compared using ImageJ.

While SIRT and EM result in high contrast and low-noise volume reconstructions, the intensity distribution show significant density inhomogeneity inside the reconstructed volume (see Fig. 1). WBP and SART result in more homogenous density inside the specimen but the noisy data makes histogram based segmentation of the unfiltered data difficult.

Besides the obvious percentile (50%) result, segmentation using Otsu and K-means gave solid‑fractions closest to the theoretical value of 50% regardless of reconstruction method (see Fig. 2).

These results show that prior knowledge about sample properties (like porosity) is useful during the segmentation step as some segmentation algorithms strongly deviate from the true structure. One should be aware that the human factor is a significant influence of on the quality and validity of tomographic reconstructions in electron tomography.

Fig. 1: YZ-orthoslices of reconstructed volumes.

Fig. 2: Table of calculated solid-fractions of reconstructed volumes.