Tooth Instance Segmentation in CBCT Images Using Watershed Algorithm and nnUNet

Achieving high-accuracy tooth instance segmentation in Cone Beam Computed Tomography (CBCT) images is critical for orthodontic preoperative planning and clinical research. However, existing methods still face significant challenges posed by tooth shape diversity, high similarity between adjacent teeth, and complex occlusal relationships. This study proposes a hybrid segmentation approach integrating the nnUNet deep learning framework and watershed algorithm. The method employs a multi-task learning network to simultaneously perform semantic segmentation of teeth and centroid localization. Subsequently, connectivity analysis and gray value assignment strategies are applied to leverage the localized centroids as seed points for the watershed algorithm, enabling precise separation of complex adjacent or overlapping tooth structures. Experimental results demonstrate that the proposed method achieves an outstanding Dice similarity coefficient of 95.94% on public CBCT tooth instance segmentation datasets, significantly outperforming existing techniques. This provides an effective solution for precise segmentation of complex tooth structures, advancing clinical applications in dental diagnostics and treatment planning. The dataset is from Cui.etl DOI: 10.1038/s41467-022-29637-2. The code is openly accessibl: https://github.com/bbsueser/MG-UnetforT.