Position-Aware Tooth Segmentation and Numbering with Prior Knowledge Injected.

Tooth segmentation and numbering are the most fundamental tasks in oral analysis as they are the prerequisite for many popular oral businesses such as root canal therapy and whitening. Despite the growing attention in this research field, the results are still far from satisfactory. Existing methods either employ a complicated multi-stage framework or couple tooth segmentation and numbering together simply by using multi-class segmentation, which is neither convenient to use nor accurate enough. To this end, we propose a single-stage multi-task framework to perform tooth segmentation and numbering in an end-to-end and decoupled fashion. Furthermore, We also involve the prior knowledge of the oral structure in the network and leverage adversarial learning to further improve the accuracy. Extensive experiments on two real-world datasets demonstrate that our proposed method achieves state-of-the-art performance.