Automatic and accurate 3D segmentation of teeth in dental cone beam computed tomography (CBCT) images is a prerequisite for computer-aided dental analysis. However, due to variations in dental anatomy, different imaging protocols, and limitations in accessing public datasets, developing an automated algorithm for dental analysis is challenging. This paper introduces a multi-stage learning-based method, named MsNet, utilizing a small amount of labeled data and a large amount of unlabeled data to achieve precise and effective 3D tooth segmentation in CBCT. In the initial stage, the nnU-Net model, incorporating both high and low-resolution modalities, undergoes training on the dataset of 12 labeled cases. Its subsequent application involves the generation of pseudo-labels for an expansive cohort of 200 unlabeled cases. By integrating the labeled and pseudo-labeled data, an optimal training dataset is constructed to train a full-resolution refined nnU-Net model for accurate and efficient 3D tooth segmentation. In two rounds of online validation, the results using MsNet are as follows: In the preliminary round, the Dice Similarity Coefficient (DSC), Intersection over Union (IoU) and Hausdorff Distance (HD) values for 3D tooth segmentation are 91.72%, 92.20% and 4.70 mm, respectively; During the final round, the values of DSC, IoU and HD are 75.08%, 80.11% and 25.55 mm, respectively.