Automatic Individual Tooth Segmentation in Cone-Beam Computed Tomography Based on Multi-Task CNN and Watershed Transform

Individual tooth segmentation in Cone-beam Computed Tomography (CBCT) is an essential procedure for the quantitative analysis of tooth function in patients with oral diseases. Though some automatic methods have been proposed with high segmentation accuracy, the overlapping tooth crown, complex shape variability of the tooth, and often inadequate signal-to-noise ratio pose serious challenges to such methods. In this study, we present a novel automatic individual tooth segmentation approach for CBCT based on multi-task convolutional neural networks (CNNs) and watershed transform. In the first stage, we utilize the multi-task CNN to segment the foreground and center region of all teeth from CBCT, in which the central regions are viewed as the landmark of the tooth for distinguishing different teeth. In the second stage, we input the predicted tooth foreground and landmark results into the marker-controlled watershed transform method to segment the overlapping tooth in the tooth foreground result by assigning each individual tooth with a unique label. In addition, we propose a 3D tooth merge strategy to improve the structural integrity of individual tooth segmentation results. The qualitative and quantitative results demonstrate that our approach achieves high accuracy and stability in terms of individual tooth segmentation for CBCT.