SegMAgNet: A Comparative Study of Segmentation Models for Defect Detection in Additively Manufactured Mg Alloys

Magnesium (Mg) alloys are increasingly used in biomedical applications due to their biodegradability, biocompatibility, and mechanical compatibility with human bone. However, fabricating these alloys using additive manufacturing techniques like Selective Laser Melting (SLM) is challenging due to their high reactivity and the potential for internal defects such as gas porosity and lack of fusion. This study presents a comprehensive evaluation of deep learning-based image segmentation methods for detecting such defects in Mg-based alloys fabricated via SLM at 80 W laser power and 500 mm/s scanning speed. X-ray Computed Tomography (XCT) was used to scan the printed samples, and 83 non-identically distributed image slices were selected to ensure variation in defect morphology. Two types of annotations—manual (using LabelImg) and threshold-based (using ImageJ)—were used to prepare datasets. These datasets were analyzed using two models: a custom U-Net with five-fold cross-validation and a U-Net with ResNet-50 backbone enhanced by patchify augmentation. The threshold-based dataset combined with U-Net + ResNet50 achieved the highest segmentation accuracy with an IoU of 86%, outperforming manual annotation-based training. This study emphasizes the importance of annotation strategy and model selection in defect detection workflows for reactive metal additive manufacturing.