The semi-supervised deep learning algorithm for effective tooth segmentation greatly simplifies the consultation process for dentists, reduces the occurrence of misclassification and missed diagnoses, and improves the efficiency of medical work. To achieve this, we propose a Tooth Boundary-aware Feature Pyramid Network (TB-FPN), a semi-supervised deep learning method applied to tooth image segmentation. This method aims to significantly improve the segmentation capability of tooth structures by integrating boundary information, effectively addressing the challenges in tooth segmentation tasks. We utilized this method on the validation set of the dental panoramic radiographs dataset using NVIDIA 3060 for tooth segmentation. The average Dice score is 91.75%, the average IoU score is 97.70%, and the average Hausdorff distance (HD) error is 0.0250. The code is available at https://github.com/Dew026/TB-FPN.git .