A tri-stream dual-branch network for three-dimensional tooth segmentation

With the advancement of three-dimensional digital dental technologies, accurate three-dimensional tooth segmentation has become increasingly important in orthodontics and computer-aided diagnosis. However, due to the variability of tooth morphology across individuals and the inherent ambiguity of tooth boundaries, achieving high-precision segmentation remains a challenging task. Existing state-of-the-art methods are primarily based on dual-stream networks utilizing coordinate and normal information, which mainly focus on local features while lacking the ability to capture global contextual structures and handle ambiguous boundary regions. To address these issues, we propose a novel artificial intelligence method based on a tri-stream dual-branch network architecture. On top of the existing coordinate and normal streams, we introduce a curvature stream, which provides stronger cues for boundary discrimination but has been treated merely as an auxiliary input in previous work. Each stream employs a dual-branch design that extracts features through both a global context branch and a local multi-scale branch. In addition, we design a novel centroid-region loss to provide auxiliary supervision during training by emphasizing representative geometric areas. Extensive experiments conducted on a real-world patient dataset demonstrate that our method outperforms existing state-of-the-art approaches in terms of segmentation accuracy. These results underscore the effectiveness of explicit geometric feature modeling and highlight its potential for advancing automated workflows in digital dentistry.