Automated tooth segmentation in Cone-Beam Computed Tomography (CBCT) images is crucial for various dental applications, including treatment planning and computer-assisted dental prosthesis design. The MICCAI STS 2024 Challenge Task 2 aims to enhance automated tooth segmentation by providing a dataset comprising both labeled and unlabeled CBCT images. This paper addresses the challenge of limited labeled data by reformulating the problem as a semi-supervised learning task. We propose a two-stage deep learning model based on nnU-Net. Our approach initially employs a low-resolution nnU-Net for quadrant segmentation, followed by a full-resolution nnU-Net for fine tooth segmentation within each quadrant. To efficiently utilize unlabeled data, we implement a selective stability-based retraining strategy to generate reliable pseudo-labels. Our method is quantitatively evaluated on the STS 2024 validation set, achieving good performance across various metrics (Dice_instance = 90.74%, Dice_image = 97.70%). The proposed approach achieved one of the highest rankings in the competition’s validation phase, demonstrating its efficacy in automatically segmenting teeth from CBCT images.