Enhancing 3D Tooth Segmentation Using Curvature-FPS Point Cloud Downsampling

3D tooth segmentation is a crucial step in dental diagnosis, treatment planning, and digital dentistry. While deep learning-based models for 3D segmentation have shown success in various real-world applications, their direct application to dentistry domain presents some challenges due to their intricate morphology of teeth and the high computational cost of processing dense 3D data. Traditional downsampling methods, such as Uniform Sampling and Farthest Point Sampling (FPS), ensure spatial coverage but often fail to preserve fine-grained anatomical structures, reducing segmentation accuracy. To address this, we propose Curvature-FPS, a novel sampling method designed for 3D tooth segmentation. By integrating curvature-based selection with FPS, our approach prioritizes key dental structures while maintaining balanced spatial distribution. We evaluate its effectiveness by training PointNet, PointNet++, DGCNN, and Point Transformer on the 3D tooth segmentation task. Our results show that the Curvature-FPS improves segmentation performance, achieving up to 4.04% higher IoU metric compared to the Uniform Sampling and 2.90% compared to vanilla FPS, with consistent gains in semantic accuracy and F1-score. Furthermore, this study contributes to the dentistry domain by systematically applying and analyzing various 3D segmentation-based models with point clouds, providing insights into their adaptation to complex dental structures.