Deep neural networks for improving physical accuracy of 2D and 3D multi-mineral segmentation of rock micro-CT images

Abstract Segmentation of 3D micro-Computed Tomographic ( μ CT) images of rock samples is essential for further Digital Rock Physics (DRP) analysis, however, conventional methods such as thresholding and watershed segmentation are susceptible to user-bias. Deep Convolutional Neural Networks (CNNs) have produced accurate pixelwise semantic (multi-category) segmentation results with natural images and μ CT rock images, however, physical accuracy is not well documented. The performance of 4 CNN architectures is tested for 2D and 3D cases in 10 configurations. Manually segmented μ CT images of Mt. Simon Sandstone guided by QEMSCANs are treated as ground truth and used as training and validation data, with a high voxelwise accuracy (over 99%) achieved. Downstream analysis is used to validate physical accuracy. The topology of each mineral is measured, the pore space absolute permeability and single/mixed wetting multiphase flow is modelled with direct simulation. These physical measures show high variance, with models that achieve 95%+ in voxelwise accuracy possessing permeabilities and connectivities orders of magnitude off. A network architecture is introduced as a hybrid fusion of U-Net and ResNet, combining short and long skip connections in a Network-in-Network configuration, which overall outperforms U-Net and ResNet variants in some minerals, while outperforming SegNet in all minerals in voxelwise and physical accuracy measures. The network architecture and the dataset volume fractions influence accuracy trade-off since sparsely occurring minerals are over-segmented by lower accuracy networks such as SegNet at the expense of under-segmenting other minerals which can be alleviated with loss weighting. This is an especially important consideration when training a physically accurate model for segmentation.