OccluDentNet : Self‐Supervised 3D Tooth Segmentation via MambaTrans Architecture

ABSTRACT In modern dentistry, digital oral scanning technology has become a key tool for acquiring precise 3D tooth models and developing treatment plans. However, efficiently and accurately segmenting teeth—especially in the absence of sufficient annotated data—remains a significant challenge. This paper introduces OccluDentNet, a self‐supervised 3D tooth segmentation method built upon the MambaTrans framework. The method partitions tooth point cloud data into multiple local patches and employs a combination of global prompting and local occlusion mechanisms to enhance the model's ability to capture both fine‐grained local details and overall shape information, thereby improving segmentation accuracy. The MambaPoint‐SA module, a key component, removes causal convolutions from the original Mamba architecture and introduces a symmetric branch to better capture inter‐point relationships, leading to improved local precision and global consistency. By integrating this module with Transformer layers, we develop a MambaTrans encoder that effectively achieves local structural modeling and global semantic awareness. Experimental results on the Teeth3DS dataset show that OccluDentNet achieves a DSC of 91.81%, which is approximately 1.6% higher than the best existing method. Its SEN is 91.65%, and PPV is 92.06%, reflecting significant improvements in both segmentation accuracy and robustness. These results demonstrate that OccluDentNet outperforms existing self‐supervised segmentation methods, with better preservation of local details and enhanced global consistency, while significantly reducing the reliance on large‐scale annotated data.