WITS: Weakly-supervised individual tooth segmentation model trained on box-level labels

Accurately and automatically segmenting teeth from cone-beam computed tomography (CBCT) images plays an essential role in dental disease diagnosis and treatment. This paper presents an automatic tooth segmentation model that combines deep learning methods and level-set approaches. The proposed model uses a deep learning method to detect each tooth’s location and size and generates prior ellipses from those detected boundary boxes. Calculating each point’s signed distance to the prior edge and using them as prior weights, the restriction term can constrain the evolution of level set functions according to the distance to the prior ellipses. Then, we use the curvature direction to find out joint points of teeth and employ a variational model to separate them to get individual results. By quantitative evaluation, we show that the proposed model can accurately segment teeth. The performance is more accurate and stable than those of classical level-set models and deep-learning models. For example, the Dice coefficient is increased by 7% than that of the U-Net model. Besides, we will release the code on https://github.com/ruicx/Individual-Tooth-Segmentation-with-Rectangle-Labels.