DGCBA: A Novel Medical Point Cloud Segmentation Network Based on Dilated Graph Convolution and Boundary Awareness

Medical point cloud analysis is a task that applies point cloud data for analysis and processing in the medical field. However, there is little research based on medical point cloud analysis. With the continuous development of graph neural networks and attention mechanisms, significant research progress has been made in the field of point cloud analysis. However, current methods focus on designing powerful feature extractors, ignoring the segmentation of geometric edge features, which in turn reduces segmentation performance. In response to the above issues, We propose a medical point cloud based segmentation network based on dilated graph convolution and boundary awareness, DGCBA. This network includes an extended geometric convolution DGConv (Dilated Graph Convolution), a multi-graph boundary-aware module (MGBA), and a learnable difference pooling module (LDP) for medical point cloud segmentation. Thus improving the accuracy of boundary segmentation. By using DGConv to construct feature relationship graphs, complex geometric edge structure information can be learned. MGBA is introduced to check global knowledge propagation on channel graphs, enrich feature representations, and LDP is introduced to obtain different global feature vectors, reducing the limitations of single pooling and fully capturing local and global feature information of point clouds. The extensive experiments of the network proposed in this article on the medical point cloud dataset IntrA and the common point cloud dataset ShapeNetPart show that our method outperforms advanced methods in segmentation tasks and also has good generalization performance.