Graph-based deep learning segmentation of EDS spectral images for automated mineral phase analysis

Graph-based deep learning segmentation of EDS spectral images for automated mineral phase analysis

Článek WoS

We introduce a novel method for graph-based segmentation of spectral images obtained using a Scanning Electron Microscope (SEM) equipped with an Energy Dispersive X-ray spectroscopy (EDS) detector. The method exploits deep learning along with fusion of rasterized electron microscopy images with sparse EDS samples to obtain accurate mineralogy segmentation with high efficiency. Improvements over previous methods are with respect to the goal of an improved quantitative and qualitative assessment of segmentation, so that volumetric composition is indirectly addressed. We describe the principles of the novel method, show experimental results on real samples and demonstrate its advantages in comparison to the state of the art. The new method performs unsupervised clustering on sparsely measured EDS spectra, allowing for classification of unseen mineralogical compounds. Then, the processed spectra are combined with single channel SEM measurements through an optimized lattice, where a Markov Field is used to perform spatial segmentation in image. The benefit of this material-agnostic method is that clusters can then be (separately) classified, analyzed, and small grains with distinct EDS measurements are more accurately separated than in previous methods. These improved results are evaluated quantitatively on ground-truth electron microscope measurements with dense high-count EDS data, as well as visually through analysis by a mineralogist.

We introduce a novel method for graph-based segmentation of spectral images obtained using a Scanning Electron Microscope (SEM) equipped with an Energy Dispersive X-ray spectroscopy (EDS) detector. The method exploits deep learning along with fusion of rasterized electron microscopy images with sparse EDS samples to obtain accurate mineralogy segmentation with high efficiency. Improvements over previous methods are with respect to the goal of an improved quantitative and qualitative assessment of segmentation, so that volumetric composition is indirectly addressed. We describe the principles of the novel method, show experimental results on real samples and demonstrate its advantages in comparison to the state of the art. The new method performs unsupervised clustering on sparsely measured EDS spectra, allowing for classification of unseen mineralogical compounds. Then, the processed spectra are combined with single channel SEM measurements through an optimized lattice, where a Markov Field is used to perform spatial segmentation in image. The benefit of this material-agnostic method is that clusters can then be (separately) classified, analyzed, and small grains with distinct EDS measurements are more accurately separated than in previous methods. These improved results are evaluated quantitatively on ground-truth electron microscope measurements with dense high-count EDS data, as well as visually through analysis by a mineralogist.

Segmentation, Deep learning, EDS spectra, Automated mineralogy

Segmentation, Deep learning, EDS spectra, Automated mineralogy

JURÁNEK, R.; VÝRAVSKÝ, J.; KOLÁŘ, M.; MOTL, D.; ZEMČÍK, P.

2023

01.08.2022

0098-3004

COMPUTERS & GEOSCIENCES

165

8

Spojené království Velké Británie a Severního Irska

1

2

2

https://www.sciencedirect.com/science/article/pii/S0098300422000668