GC-UNet: Enhance UNet with GCN for Periodontitis Segmentation

The intelligent segmentation of periodontitis regions plays a crucial role in the field of dental medicine, significantly aiding in the diagnosis and treatment of the disease. In the domain of medical image segmentation, the UNet architecture, composed of convolutional networks (CNNs), has shown outstanding performance, yielding favorable results in many segmentation tasks. However, when dealing with periodontitis segmentation images that feature both complex overall structures and intricate details with numerous focal points, a UNet based on CNNs suffers from insufficient receptive fields and inadequate capability to capture global feature relationships. Therefore, in this paper, we propose GC-UNet, which designs a GCN-based global feature receptor. This receptor embeds the feature map information obtained by the UNet encoder into graph data and employs a method to calculate edge weights based on spatial distance. By cleverly converting the original spatial positional relationships into an adjacency matrix, our network can capture the associations of global information. To this end, we design a series of experiments to compare the performance of our method with existing methods on the periodontitis segmentation task and to investigate the impact of various key parameter variables on network performance. The experiments demonstrate that GC-UNet exhibits superiority over existing methods in the periodontitis segmentation task.