Multi-mineral segmentation of micro-tomographic images using a convolutional neural network

Abstract

Micro X-ray-Computed Tomographic (micro-CT) images are often used for calibration of interpretation models that relate physical rock properties to the rock microstructure. Some computed properties, for example, permeability, may be relatively insensitive to the mineral composition of the rock matrix, while electrical conductivity and stiffness show strong dependence on the volume fractions and spatial distribution of minerals. This information could potentially be extracted from X-ray density of these images through a procedure commonly called segmentation. However, the range of X-ray density overlaps for different minerals, hence an effective segmentation workflow need to consider both the density values and shapes of the minerals. Such image processing workflows consist of many stages, and thus become time and computationally expensive and involve a lot of manual labor. This paper proposes an automated workflow for the multi-mineral segmentation of micro X-ray-Computed Tomographic (micro-CT) images using a convolutional neural network (CNN). The CNN model is trained using labels of two sets of images of a Bentheimer sandstone that are segmented into pore, quartz, clay and feldspar using a sophisticated interactive workflow. The trained model is then used to segment a new set of images of the Bentheimer sandstone. The segmented multi-mineral labels can achieve an accuracy of ∼97% and the process takes only ∼10 min as compared with interactive workflow which takes ∼3 h. Although, CNN-based segmentation algorithms were published in the literature before, the proposed model is capable of more sophisticated segmentation and achieves superior accuracy on a completely blind test set. Potentially, our approach might generalize well to other lithology for the micro-CT image analysis.