Porosity segmentation in X-ray computed tomography scans of metal additively manufactured specimens with deep learning

Selective laser melting (SLM), a key technology in metal AM, is extensively applied in industries such as aerospace, automotive, and medical sectors. SLM process parameters and different types of defects that may appear during the manufacturing process affect the quality of the final product. X-ray computed tomography (XCT) enables non-destructive characterization of internal features while mitigating interference from complex geometries and external noise. However, accurate pixel-level segmentation remains challenging due to weak contrast, blurred boundaries, and extreme class imbalance of tiny pores. This paper proposes a semantic segmentation model based on an improved encoder-decoder architecture. The model integrates residual modules to stabilize deep feature learning and alleviate gradient degradation, and a spatial–channel attention mechanism that extracts cues from the deepest encoder features and generates scale-specific parameters to modulate decoder reconstruction, thereby emphasizing pore boundaries while suppressing redundant channels. The proposed model is validated on a public CoCr XCT dataset and a multi-defect Inconel 718 dataset established in this study. On the CoCr dataset, our method achieves 78.94% MIoU and 98.92% ACC, outperforming U-Net by +3.53% MIoU. On the Inconel 718 dataset, the model achieves 95.74% and 82.03% segmentation accuracy for keyhole and LOF defects, respectively, with an average accuracy of 93.22%, exceeding U-Net, FCN and PSPNet.