Multi-Encoding Contrastive Learning for Dual-Stream Self-Supervised 3D Dental Segmentation Network

To address the limitation regarding the supervised dataset scale in the semantic recognition of newly distributed types such as wisdom teeth and missing teeth, the multi-encoding contrastive learning for dual-stream self-supervised 3D dental segmentation network (MECSegNet) is proposed. First, a self-supervised encoder pre-training framework is designed by integrating 3D mesh feature representation to construct a deep feature encoding network, where the pre-trained encoder learns universal dental feature representations. Then, a multi-contrastive loss function is established to jointly optimize the self-supervised encoder, extracting effective local and global feature representations while incorporating a cross-stream contrastive loss to learn discriminative features from multiple perspectives. Finally, the improved encoder is integrated into a dual-stream network to build a fine-tuning framework for supervised fine-tuning on a small proportion of labeled data. Experimental results show that, with only 20% labeled data, the proposed MECSegNet achieves a 1.3% improvement in accuracy and a 79.81% reduction in computational cost compared to existing self-supervised methods, while maintaining comparable segmentation accuracy and efficiency to high-performance fully supervised methods.