Ahmed Musa1,2, Baeckkyoung Sung1,3*, Leon Abelmann4
1Biosensor Group, KIST Europe Forschungsgesellschaft mbH, 66123 Saarbrücken, Germany
2Department of Computer Science, Saarland University, 66123 Saarbrücken, Germany
3Division of Energy & Environment Technology, University of Science & Technology (UST), Daejeon 34113, South Korea
4Faculty of Electrical Engineering, Mathematics & Computer Science, Delft University of Technology, 2628 CD Delft, the Netherlands
sung [at] kist-europe.de
Abstract
Characterisation of organic and inorganic microparticles has long been an important topic in the field of environmental health sciences. Especially, combined analytical method of scanning electron microscopy (SEM) associated with energy-dispersive X-ray spectroscopy (EDX) is a commonly exploited approach to obtain extensive data on the size, morphological, and elemental information from the particulate specimens. Particulate matter (PM) is a representative atmospheric pollutant that may exert adverse effects on the human respiratory system, and SEM-EDX is a widely used tool for extracting PM analysis data, which can be subsequently utilised as physicochemical features for toxicological predictions.
In this presentation, we show a machine learning-based automation of SEM-EDX correlation of environmental PM data. First, we segment SEM images using WEKA trainable segmentation which is based on a random forest algorithm to classify pixels as foreground and background groups, followed by finding connected components (pixels that are foreground and connected vertically or horizontally). These regions are used to calculate PM shape parameters. Next, element maps are obtained from EDX using curve fitting with HyperSpy Python package. PM regions from SEM images are utilised to sum intensities in the same spatial location for the element maps to obtain elemental profiles. We finally build two models to predict PM elements: (1) Element maps from SEM-EDX data using image-to-image translation, and (2) regression to predict PM element percentages from shape features. Results from model 1 and 2 are then applied to extract PM elemental profiles associated with PM morphology information. Our results show how to efficiently predict EDX and element maps from SEM images with a high degree of accuracy. This method has a potential to significantly reduce time and labour required for environmental PM monitoring.
Keywords: Environmental health, particulate matter, SEM, EDX, automated data analysis, multiple output regression
