Medical image segmentation is a critical task for accurately delineating target and pathological regions. Despite significant progress in deep learning, existing methods still struggle to integrate local features with global contextual information, especially in tooth CBCT images with complex and irregular morphology. To address these challenges, we propose LGF-Net, a dual-branch encoder–decoder network that fuses local and global features to enhance segmentation accuracy and robustness. In the encoder, multiple DSsm modules combine Mamba mechanisms with dense connections, with each Local–Global Block (LG-Block) extracting local details and global context in parallel before feature fusion. An Adaptive Cross-scale Feature Fusion (ACFF) module at the bottleneck promotes multi-level feature interaction, while the decoder progressively restores spatial resolution through upsampling and skip connections. Experiments on two widely used CBCT datasets demonstrate that LGF-Net achieves state-of-the-art performance, significantly outperforming mainstream methods in DSC and IoU, and maintains high accuracy even in the presence of metallic artifacts, highlighting its robustness and clinical potential.