Methane (CH4) hydrate dissociation and CH4 release are potential geohazards currently investigated using X-ray computed tomography (XCT) imaging in laboratory experiments. Image segmentation constitutes an important data processing step for this type of research, but it is often time consuming, computing resource-intensive and operator-dependent. Furthermore, segmentation procedures are frequently tailored for each XCT dataset due to differences in image characteristics, such as greyscale contrast variations. To address these issues, an investigation has been carried out using U-Nets, a class of Convolutional Neural Network, to segment synchrotron radiation XCT (SRXCT) images of CH4-bearing sand during hydrate formation. Emphasis was given to CH4 gas bubbles, due to their paucity and low contrast. Three U-Net deployments previously untried for this task were assessed: (1) a bespoke 3D hierarchical method, (2) a 2D multi-label, multi-axis method and (3) RootPainter, an application that combines a 2D U-Net with interactive corrections. U-Nets were trained using very small hand-annotated datasets to reduce operator time. Results show high segmentation accuracy and consistency, with RootPainter slightly outperforming the alternative approaches and all three methods surpassing mainstream watershed and thresholding techniques. Greyscale contrast between material phases was found to affect segmentation performance, with the lowest metrics corresponding to data exhibiting the lowest contrast. Segmentation accuracy affected derived parameters such as CH4-saturation and porosity, but errors were small compared with gravimetric methods. It was also found that U-Net models trained on low greyscale contrast images could be used to segment higher-contrast datasets and also data collected at a different facility, thereby demonstrating model portability. Such portability is anticipated to be advantageous when the segmentation of large XCT datasets needs to be delivered over short timespans.