MNFF-GNN: Multi-Order Neighbor Feature Fusion Graph Neural Network

With the rapid development of artificial intelligence, automated artifact recognition technology has gradually replaced the traditional manual quality evaluation method. The existing samples of CT images containing artifacts are small, and the relationships between the images are of great significance. In this study, firstly, a method for CT image artifact recognition was developed by transforming the problem into a node classification framework. Secondly, the characteristics of this complex network and the features of the CT image texture were extracted. Finally, the combination of the complex network's characteristics and CT image texture features was viewed as node attribution; the relationship between different nodes was analyzed using a graph attention network; and classification was carried out. The integration of multi-order neighbor features in the MNFF-GNN model improves the representation of motion artifact regions, targeting the limitations of traditional methods and convolutional neural networks (CNNs). The model demonstrates potential as a clinical tool, particularly in resource-constrained settings, by effectively identifying artifacts even with limited data, with an accuracy of 90.9%, which is an improvement of 9.73%. This innovative approach leverages graph neural networks (GNNs), which are particularly effective at capturing both local and global relationships within graph-structured data.