Manual segmentation of 3D and 4D synchrotron X-ray computed tomography (sXCT) data is tedious, time-consuming, and often requires human experts with a high level of domain knowledge. Deep learning tremendously facilitates and aids in this task. Convolutional neural networks (CNNs) trained on labelled data have become state-of-the-art for image segmentation tasks. However, these black-box models lack explainability and tend to be overly confident in their predictions. Moreover, under-represented classes or lower-dimensional structures are often hard to predict correctly.

In this study, we implemented several computation methods allowing explainability and uncertainty estimations. The uncertainty estimation highlights areas where the model is truly unsure about its prediction. This opens the possibility to segment under-represented classes. For instance, areas of high uncertainty correlate well with rarely occurring cracks in our data set. Additionally, heatmap-based explainability methods like Grad-CAM show the potential to improve model selection and reveal weaknesses of trained models. Overall, we demonstrate how uncertainty and explainability methods improve CNN-based segmentations of sXCT data.