ANTEMA: a neural network-based automated nanoparticle transmission electron micrograph analysis

Abstract text (incl. figure legends and references)

High-resolution transmission electron microscopy (HRTEM) is an important tool for nanoparticle analysis. This is because HRTEM images show information about the size and shape of the particles as well as their inner structure. These are all parameters that influence the physical and chemical properties of the particle and are therefore of importance for all kinds of nanoparticle applications. With high frame rates and increasing capabilities of transmission electron microscopes, large amounts data are generated. However, analysing these micrographs is often done manually which is a tedious and time-consuming process. This is especially important for in-situ TEM experiments where hundreds or thousands of frames must be analysed. As a result, only a fraction of the available data is usually quantitively analysed.

To facilitate faster and more in-depth analysis of HRTEM images we have developed an automated image processing module using machine learning techniques. An important step in realizing an automated analysis of the nanoparticles in HRTEM images is to separate the particles from the noisy amorphous background. This is realized by a neural network which was trained on labelled HRTEM images by a supervised learning approach. The segmentation maps that are generated by the network are further processed to exclude particles that cross the image boundary and identify and split overlapping particles. From the resulting binary maps that include all particles in the image, shape and size-related parameters like circularity, equivalent diameter, and Feret diameter are extracted. These steps are all integrated in an automated routine which is able to analyse an HRTEM image within a few seconds. This permits the automated analysis of large quantities of data and gives more insight into nanoparticle structures compared to a manual analysis. It also avoids an operator-induced bias during manual image analysis.