Application of machine learning techniques in mineral phase segmentation for X-ray microcomputed tomography (µCT) data

Highlights

• Unsupervised classification for rapid segmentation between gangue and sulfides.

• Feature matching to align BSE image and µCT slice for creating training data.

• Supervised classification correctly classifies 60% of chalcopyrite and 99% of pyrite.

• Feature-based classification produces better segmentation between mineral grains.

• Overlapping grayscale values in the µCT data limits the classification accuracy.

Abstract

X-ray microcomputed tomography (µCT) offers a non-destructive three-dimensional analysis of ores but its application in mineralogical analysis and mineral segmentation is relatively limited. In this study, the application of machine learning techniques for segmenting mineral phases in a µCT dataset is presented. Various techniques were implemented, including unsupervised classification as well as grayscale-based and feature-based supervised classification. A feature matching method was used to register the back-scattered electron (BSE) mineral map to its corresponding µCT slice, allowing automatic annotation of minerals in the µCT slice to create training data for the classifiers. Unsupervised classification produced satisfactory results in terms of segmenting between amphibole, plagioclase, and sulfide phases. However, the technique was not able to differentiate between sulfide phases in the case of chalcopyrite and pyrite. Using supervised classification, around 50–60% of the chalcopyrite and 97–99% of pyrite were correctly identified. Feature based classification was found to have a poorer sensitivity to chalcopyrite, but produced a better result in segmenting between the mineral grains, as it operates based on voxel regions instead of individual voxels. The mineralogical results from the 3D µCT data showed considerable difference compared to the BSE mineral map, indicating stereological error exhibited in the latter analysis. The main limitation of this approach lies in the dataset itself, in which there was a significant overlap in grayscale values between chalcopyrite and pyrite, therefore highly limiting the classifier accuracy.

Introduction

There has been growing interest in X-ray microcomputed tomography (µCT) application in geosciences, due to its non-destructive nature that allows three-dimensional (3D) analysis of an object. µCT could potentially eliminate stereological errors generated by conventional two-dimensional microscopy analysis used for ore and rock samples, allowing more accurate analysis of the samples. Rapid development of µCT systems currently allows spatial resolution down to nanometer scale (Ghorbani et al., 2011, Yang et al., 2017), as well as enabling in-situ experiments to be performed during acquisition, thereby acquiring time-based 3D data (Ghorbani et al., 2011, Lin et al., 2016a, Lin et al., 2016b). These developments make µCT a more attractive analytical technique for rock samples.

Various applications of µCT systems in geoscience have been studied, including mineral liberation and grain analysis, pore and fracture analysis, and to some degree, texture analysis (Garcia et al., 2009, Ghorbani et al., 2011, Jardine et al., 2018, Lin and Miller, 2005, Lin and Miller, 1996). Reviews discussing current and potential applications of µCT system in geosciences have also been published (Cnudde and Boone, 2013, Guntoro et al., 2019, Kyle and Ketcham, 2015, Maire and Withers, 2014, Mees et al., 2003, Miller et al., 1990). Nevertheless, the current use of µCT for ore samples is more focused towards structural analysis of the samples, such as analysis of pores and fractures (Deng et al., 2016, Ghorbani et al., 2011, Müter et al., 2012). Other uses include particle size distribution analysis (Wightman et al., 2015) and particle shape analysis (Lin and Miller, 2005). In a recent study, textural analysis of mineral phases in a drill core sample was conducted using a µCT system through the correlation and association indices between volume elements (voxels) in the 3D dataset (Jardine et al., 2018).

Being able to use µCT to generate a 3D structural analysis as well as gaining mineralogical information of the ore sample at the same time would add value to the technique as well as its potential uses. This is where µCT is currently limited due to challenges including: similar attenuations between mineral phases, limited resolution, and lack of automated mineralogical analysis software. Both pre- and post-scanning techniques aiming to obtain a reliable 3D mineralogical analysis have been evaluated by several researchers (Bam et al., 2019, Ghorbani et al., 2011, Kyle et al., 2008, Reyes et al., 2017, Tiu, 2017, Wang et al., 2015).

Pre-scanning techniques refer mostly to optimization of the scanning conditions as well as calibration with pure minerals. Scanning conditions can be adjusted to increase the attenuation contrasts between minerals, which often means using lower voltage and/or reducing sample size. Reyes et al. (2017) have shown that segmentation between chalcopyrite and pyrite was still found to be difficult when using voltage as low as 50 kV. Kyle et al. (2008) have shown that bornite, chalcopyrite, and magnetite minerals could be differentiated with smaller-diameter (≤22 mm) core samples at scanning energy of 180 KeV. Additionally, using smaller sample size could also suppress the beam hardening effect (Bam et al., 2019), which can contribute to segmentation inaccuracies (Reyes et al., 2017).

Calibration with high purity mineral samples in combination with dual energy µCT scanning has been demonstrated by Ghorbani et al. (2011). By calibrating the µCT system with minerals with known density, a correlation that relates density with the attenuation coefficient can be obtained. The density of the material can also be determined directly through the relation between attenuation coefficients obtained from scanning at two different energy levels. Ghorbani et al. (2011) combined both procedures so that the measured density from dual energy scanning (130 and 200 kV) can be compared against the real density of the calibration samples. Using such procedures, Ghorbani et al. (2011) successfully differentiated pyrite, quartz, and sphalerite minerals in the sample.

Post-scanning techniques refer to the image processing procedures applied to the acquired µCT dataset. If the differences in attenuations are significant enough, simple thresholding techniques such as the one developed by Otsu (1979) can be used to set a threshold between the grayscale values and subsequently differentiate the phases. This method has been used in segmenting pores/air and the mineral matrix, as well as heavy and light minerals (Andrä et al., 2013, Lin et al., 2016a, Lin et al., 2015, Reyes et al., 2017, Yang et al., 2017). In cases where the attenuation differences are insignificant, cross-correlation of the µCT dataset to other dataset such as dataset obtained from Scanning Electron Microscope with Energy Dispersive X-ray Spectroscopy (SEM-EDS) has been shown to be capable of distinguishing minerals with similar attenuations such as copper sulfides and pyrite (Reyes et al., 2017, Tiu, 2017).

Another promising approach in differentiating between mineral phases is the use of machine learning techniques. In general, machine learning is defined as the use of mathematical models to interpret the underlying patterns in a dataset. By learning this pattern, a computer system can make predictions or classifications on the dataset (Suthaharan, 2016). Machine learning can be divided into unsupervised and supervised learning. Supervised learning means that the user pre-defines the underlying pattern of the data, and the computer builds a prediction model based on the pre-defined pattern (training data). Unsupervised learning lets the computer interpret the pattern by itself without user’s supervision.

Several recent studies have evaluated the use of machine learning in the segmentation of mineral phases in µCT dataset. Chauhan et al., 2016a, Chauhan et al., 2016b) evaluated the performances of various machine learning algorithm for µCT datasets, focusing mostly on the segmentation of pores from the rock matrix. Both unsupervised and supervised algorithms were evaluated, including K-means, Fuzzy C-means, Self- Organized Maps (SOM), Artificial Neural Network (ANN), as well Support Vector Machines (SVM). Tiu (2017), evaluated the use of supervised classification (random forest) in segmentation between chalcopyrite and pyrite in a drill core sample, using SEM-EDS mineral map as training data. Wang et al. (2015), used feature-based random forest classification to segment multiphase particulate samples from the background, in which they concluded that the resulting segmentation from simple thresholding technique was not satisfactory.

This study systematically evaluates the application of different machine learning techniques in mineral segmentation to a µCT dataset. Both unsupervised and supervised learning techniques are included in this study. Additionally, an automated image registration technique is introduced to align a Back Scattered Electron (BSE) mineral map with a corresponding slice in a 3D µCT data, which is then used as the training data to classify the other µCT slices. Furthermore, besides using grayscale values as the dataset, the possibility of classification using features such as edges, corners, and blobs (regions with similar grayscale values) is also evaluated. The accuracy and computational costs of these methods are evaluated and compared to give insight on the most suitable method for specific tasks related to mineral segmentation of a µCT dataset.