Machine-learning approaches for the enhancement of 2D and 3D electron microscopy data

Abstract

Electron microscopy allows academic and industrial users to investigate the structure and properties of a variety of materials. However, some limitations are imposed by the instrumentation and the nature of the analyzed specimen. This work proposes machine-learning-based approaches to overcome some of nowadays impediments in the field of 2D and 3D electron microscopy data. In the first project, machine learning is employed to enhance the quality of Scanning Transmission Electron Microscope (STEM) data, effectively reducing noise levels across various electron beam intensities. The algorithm developed undergoes rigorous testing using both synthetic and actual microscopy data. Furthermore, a quantitative and impartial benchmarking protocol for comparing various denoising workflows is proposed, based on the precision of atomic column localization. The second project focuses on STEM data analysis in the context of quantifying vacancies in transition metal dichalcogenides (TMD). Here machine learning improves the quality of STEM-acquired TMD images, facilitating the vacancy-counting process in materials science research. The third project explores the application of a powerful neural network, developed for video-frame interpolation, for the enhancement of 3D tomography. This innovative approach significantly increases the resolution of tomographic images, with applications ranging from materials science, where it aids the study of graphene nanosheets, to medical imaging, where it potentially reduces ionizing radiation doses in Computed Tomography (CT) scans and enhances cardiovascular assessment in coronary angiography videos. Throughout the entire work, a particular effort is dedicated to the development of the techniques needed to quantify the improvement resulting from the application of the proposed methodologies, which are compared to the state-of-the-art approaches. This research development demonstrates the versatility and transformative potential of machine learning in advancing imaging techniques across diverse scientific domains.