Deep Learning-Based Defect Segmentation in 3D Tomography of Castings: A Complete Pipeline with Annotation, Model Comparison, and Object-Level Evaluation

X-ray computed tomography (CT) enables accurate three-dimensional characterization of internal defects in aluminum castings. However, the manual delineation of gas porosity and shrinkage defects across large CT volumes is time-consuming and subject to operator-dependent variability. Deep learning provides a promising approach for automating this task, but its development remains limited by the availability of accurately annotated volumetric data. This study addresses these challenges through three complementary contributions. First, a semi-automatic annotation tool was developed to facilitate the generation of reference segmentation masks. Second, three model configurations were evaluated: a fully supervised 3D U-Net, a fully supervised 3D Attention U-Net, and a 3D U-Net pretrained through self-supervised volumetric inpainting and subsequently fine-tuned for defect segmentation. Third, the models were evaluated using both voxel-level metrics and object-level analysis to assess segmentation accuracy and practical defect-detection performance while accounting for fragmented predictions. For model development, 80 representative annotated CT subvolumes of aluminum castings were used for training, while an independent set of 20 CT subvolumes was reserved for final evaluation. With volume weighting and the specified object-level detection criterion, the inpainting-pretrained 3D U-Net achieved mean precision, recall, and F1-score of approximately 91%, 96%, and 93%, respectively. Its mean voxel-level Dice coefficient was approximately 65%. These results highlight the potential of volumetric self-supervised pretraining to improve the robustness and reliability of deep-learning-based defect inspection in X-ray CT images of aluminum castings.