<p>Dental segmentation plays a vital role in 3D computer-aided orthodontics, yet achieving precise segmentation results remains a challenge due to the unique shapes of teeth, their varying arrangements, and prevalent issues such as missing or crowded teeth, as well as the limited availability of annotated data. Existing approaches often require extensive manual interaction or suffer from inaccuracies. In response, we present a novel and adaptable Attention U-Net network tailored explicitly for segmenting 3D dental models. To address data limitations and enhance model robustness, our method incorporates a preprocessing stage based on the Multiresolution Adaptive Parameterization of Surfaces algorithm, which augments the dataset by introducing morphological variations while preserving overall anatomical structure. Within the neural architecture, our approach leverages the conceptual similarity between mesh faces and 2D image pixels to adapt convolution, pooling, and up-sampling operations in the Attention U-Net architecture. Furthermore, our approach incorporates an efficient post-processing step to correct incorrectly predicted segments within teeth, resulting in heightened segmentation performance. For the purpose of evaluating our proposed approach, we performed experiments on a 3D dental dataset collected by expert clinicians. Comparative analysis against state-of-the-art methods demonstrated that our approach achieved superior accuracy in dental segmentation tasks.</p>