Aircraft icing poses a serious threat to flight safety, and accurately identifying the three-dimensional (3D) pore structures within dynamic ice is crucial for informing anti-icing and deicing strategies. Although micro-nano computed tomography (CT) offers a nondestructive means of capturing the internal morphology of ice, segmenting the pores from CT volumes remains a significant challenge due to irregular geometries, weak boundaries, and limited labeled data. To overcome these challenges, we propose a semisupervised 3D segmentation framework named StructGNN-UNet, which integrates a 3D UNet backbone with a graph neural network (GNN) module for capturing spatial structural relationships between voxels. In addition, we introduce a structure-aware loss function that combines supervised binary cross-entropy (BCE), edge smoothing loss, structural consistency loss, and pseudolabel loss to enhance both global and local segmentation consistency, especially across blockwise reconstructions. Extensive experiments conducted on real CT datasets of aircraft wing icing demonstrate that our method significantly reduces pore disconnection at block boundaries and improves segmentation within blocks. Compared to the standard mean teacher, StructGNN-UNet achieves a higher F1 score of 98% and an accuracy of 97%, surpassing the baseline by 5% and 6%, respectively, while achieving performance comparable to training with fully labeled data. The proposed StructGNN-UNet framework effectively addresses the limitations of traditional blockwise segmentation strategies. By incorporating spatial structure modeling and structure-aware learning, the proposed method provides a robust and generalizable solution for segmenting complex porous structures formed during dynamic icing on aircraft surfaces, supporting more accurate analysis of icing patterns and contributing to improved flight safety and deicing strategies in aerospace applications.
Tang et al. (Sun,) studied this question.