Point cloud applications in medical images

This thesis comprehensively investigates kidney and pancreas segmentation problems using voxel-based and point cloud–based deep learning methods. The study aims to evaluate the impact of data representation on model performance and generalization capability. For kidney segmentation, two-stage coarse-to-fine framework was designed, and architectures including U-Net, AttUNet, ResUNet, MFFAU-Net, and TransUNet were tested on KiTS19, KiTS23, and FLARE22 datasets. The results demonstrated that voxel-based methods achieve high accuracy but remain limited by computational cost and generalization challenges. To address these limitations, point cloud–based architectures, PointNet and DGCNN, were explored. These models achieved competitive results, benefiting from low memory usage and strong geometric awareness, particularly in delineating complex organ boundaries. The DGCNN model achieved superior accuracy compared to voxel-based methods in pancreas segmentation. Overall, this study systematically demonstrates the applicability of point cloud–based representations for medical image segmentation and provides novel insights into the relationship between data representation and model performance, offering valuable guidance for future research.