Optimized segmentation for quantitative 3D analysis of electron tomography reconstructions

Abstract text (incl. figure legends and references)

Nanoparticle assemblies are of great interest due to their tunable structure and properties, and their application in various fields.1,2 For example, the coexistence of Au nanoparticle and quantum dots (QDs) enables dual imaging, fluorescence and light scattering, and/or photothermal treatment of cancer cells.3 By changing the number of particles, and/or the interparticle distances, it becomes possible to reach specifically desired structures and properties for targeted usage.4 Hereby, it is very important to investigate the structure-property relationship of the nanoparticle assemblies. In order to extract quantitative information of the 3D structure, an accurate characterization at the nanometer level is required. Transmission electron microscopy (TEM) is a useful technique to study nanomaterials. However, since conventional TEM is limited to 2D projections, it is important to perform electron tomography experiments to obtain reliable information for 3D structures. An important step to extract quantitative 3D information is the segmentation of the 3D data set. Manual segmentation is hereby the most conventional but also most time consuming technique. To overcome this limitations, advanced reconstruction algorithms, such as the discrete algebraic reconstruction technique (DART),5 and the sparse sphere reconstruction algorithm (SSR) 6 have been developed in which segmentation is part of the 3D reconstruction. However, they are computationally expensive and might not be ideal for assemblies where the nanoparticles are agglomerated or not uniform in size. In this work, we therefore focused on the optimization of the segmentation process and we have compared three different segmentation methods to extract quantitative parameters from the 3D reconstructions. In additional to the manual segmentation (Figure 1), we made use of a watershed function7, which separates the agglomerated particles from their catchment points for each local minima, and a spherical Hough function, which detects each particle separately by defining their radius and center point.8 Results showed that the spherical Hough function, is the most promising technique to extract quantitative information from the 3D reconstructions of the nanoassemblies. In this contribution, I will present the results and comparison of these segmentation methods, including their advantages and limitations.

1. Hanske, C. et al. M. Advanced Materials 30, 1707003 (2018).

2. Colloidal Synthesis of Plasmonic Nanometals. (2020).

3. Jin, Y. & Gao, X. Nature Nanotechnology 4, 571–576 (2009).

4. Hollingsworth, J. A. et al. MRS Bulletin 40, 768–776 (2015).

5. Batenburg, K. J. et al. Ultramicroscopy 109, 730–740 (2009).

6. Zanaga, D. et al. Nanoscale 8, 292–299 (2016).

7. Beucher, S. & Meyer, F. in Mathematical Morphology in Image Processing (CRC Press, 1993).

8. Hough, P. V. C. Method and Means for Recognizing Complex Patterns. (1962).

Figure 1: The samples: CdSe/CdS quantum dots encapsulated by polymeric micelles (a,e), Au nanosphere and CDSe/CdS quantum dots encapsulated by polymeric micelles (b,f), Au nanorod and CdSe/CdS quantum dots assembled on silica shell (c,g), and Au nanorod and PbS quantum dots assembled on silica shell (d,h). The first row (a-d); 3D visualizations of the reconstructions of nanoassemblies, and the second row (e-h); manual segmentation of the same structures. (Scale bars represent 30 nm).