In the realm of dental imaging, the utilisation of 3D dental cone-beam computed tomography (CBCT) scans has gained significant prominence, particularly in the fields of orthodontics and endodontics. These scans, characterized by their lower radiation exposure, have proven instrumental in clinical diagnosis. A crucial aspect of their application involves the precise segmentation of teeth, as it provides vital anatomical information for medical practitioners and serves as a cornerstone in computer-aided diagnosis. While the nnU-Net architecture has gained popularity for medical image segmentation, the practical integration of unlabelled medical image data remains a significant challenge. In response to this, our paper introduces a semi-supervised approach rooted in the foundational nnU-Net structure. This method incorporates the preprocessing of data with prior knowledge and iterative optimisation of pseudo-labels to enhance the accuracy of tooth segmentation.A key insight from our research is the pivotal role played by morphological opening and closing operations as preprocessing steps, capitalizing on the wealth of a prior information available. This strategic approach culminated in remarkable results, with a preliminary-round score of 0.9381 and a final-round score of 0.8427, demonstrating the effectiveness of our proposed technique.