In the field of medical imaging, tooth segmentation based on 3D Cone-Beam Computed Tomography (CBCT) is recognized as a very challenging task. Precise segmentation of the teeth is crucial for dental diagnosis and treatment planning, providing dentists with detailed tooth structure information to facilitate personalized treatment planning and improve the success rate of clinical treatment. Based on nnUNet, we developed a tooth segmentation method suitable for 3D CBCT data. This innovative training process combines a semi-supervised learning method based on Kullback-Leibler divergence and a supervised learning strategy combined with dynamic convolution and introduces morphology-based preprocessing operations in data processing. In the MICCAI STS 2023 Challenge: STS-3D CBCT-based tooth segmentation task, our method achieved a Dice similarity coefficient of 0.9111 and 0.7261, an IoU of 0.9164 and 0.7855, and a 3D Hausdorff distance of 0.0453 and 0.2595 on the preliminary and rematch test data set.