Additive manufacturing (AM) is increasingly gaining interest as a low-waste production technique, capable of producing objects using a computer-aided design file. It is particularly interesting for rapid prototyping of parts and manufacturing objects that have complex shapes. However, as in the case of AM through selective laser melting (SLM), manufactured objects may contain defects that can seriously alter their properties. These defects may appear as pores, which can be detected by X-ray computed tomography (X-CT) in a non-destructive manner. CT images can simply be segmented by thresholding or through more advanced techniques such as discrete X-CT reconstruction or machine learning techniques. Nevertheless, these techniques are vulnerable to image reconstruction artefacts. In this work, we evaluate the performance of state-of-the-art, deeply supervised 3D deep learning networks (UNet++, UNet 3+ and UNet-MSS) in terms of segmentation performance of pores from X-ray CT images. The networks have been trained on a real CT dataset, with (noisy) labels produced from both conventional thresholding of the CT images as well as more advanced discrete polychromatic reconstructions. Furthermore, the performance of the networks was evaluated on a test dataset with severe CT artefacts. Pore segmentation from real CT images, which include noise and reconstruction artefacts, revealed that the best performing network was UNet++ with an average Sørensen-Dice score of 0.869 ± 0.006.